A method and system for assisting diagnosis of dysplasia from whole-slide images

Abstract

The invention provides a computer-implemented method for assisting diagnosis of dysplasia from a liquid-based cytology whole-slide image (WSI), using a multi-stage pipeline combining deep- learning-based detection with engineered feature extraction and a calibrated machine-learning classifier. The method involves receiving and resampling a digital WSI to a canonical spatial resolution, detecting individual cell instances with a neural network, computing morphological and texture features for each detected cell, aggregating these features across the WSI to form a slide-level feature vector, inputting this vector into a calibrated classifier to yield a slide-level probability of dysplasia, computing slide-quality metrics to govern downstream triage, presenting results via an interactive graphical interface with per-cell thumbnails, and iteratively refining the system with expanded and harmonized training data from expert annotations.

Description

[0001] A METHOD AND SYSTEM FOR ASSISTING DIAGNOSIS OF DYSPLASIA FROM

[0002] WHOLE-SLIDE IMAGES

[0003] TECHNICAL FIELD

[0004] The present invention relates to methods and systems for the diagnostic analysis of cytology microscopic images, such as Papanicolaou (PAP) smears, HE / Giemsa, using whole-slide imaging (WSI).

[0005] BACKGROUND

[0006] Digital pathology is an evolving field that uses advanced imaging and computational techniques to analyze tissue samples. This approach has transformed traditional pathology by enabling high-resolution digital scans of entire slides, known as whole-slide images (WSIs). These digital slides can be stored, shared, and processed with algorithms to support diagnostic work, workload management, and quality assurance.

[0007] One key application is the analysis of cytology microscopic images, such as Papanicolaou (PAP) smears, for the detection of cervical dysplasia and cervical cancer. Traditionally, pathologists examine PAP smears under a microscope to identify abnormal cells. This manual process is labor-intensive, time-consuming and subject to inter- and intra-ob server variability.

[0008] Whole-slide imaging involves scanning an entire glass slide at high resolution, often producing images measuring tens or hundreds of thousands of pixels per dimension and reaching gigabyte-scale file sizes. The size and resolution of WSIs pose practical challenges for computation, memory and storage. Common strategies such as dividing slides into tiles reduce peak memory but introduce additional issues, including overlap handling and consistent mapping between tile and slide coordinates.

[0009] Beyond scale, input heterogeneity hinders consistent analysis. WSIs differ by scanner optics and acquisition settings, resulting in varying micrometers-per-pixel (MPP) values, even for nominally similar magnifications. Stain appearance also varies across laboratories and over time. In practice, downstream measurements (e.g., areas, distances, densities) may be reported in raw pixel units or site-specific scales, complicating cross-scanner comparability and reproducibility when metadata are missing or unreliable.

[0010] Image quality itself is variable. Cytology WSIs can exhibit regions with suboptimal brightness, color balance, contrast or focus, as well as artefacts such as debris, air bubbles and precipitate. Quality fluctuations occur at the slide level and locally within tiles. Although various metrics (e.g., brightness, colorfulness, blur or edge-based focus measures) are known, consistent policies for turning metric values into operational decisions (e.g., caution vs. block) and for linking those decisions to triage are not universally implemented.

[0011] Cytology also includes liquid-based cytology (LBC) preparations (e.g., ThinPrep or SurePath), where cells are distributed in monolayers with between-slide variability in cellularity, debris / mucus burden and stain hue. LBC platforms differ in layer thickness and focus spread, which can affect both quality metrics and detection behavior. Many published approaches do not explicitly account for platform-specific quality bands or for cross-site harmonization in LBC workflows.

[0012] Modern computer-vision systems, including object detectors, have been applied to cytology for pre-screening and candidate cell localization. However, existing approaches often rely on detector scores alone, heuristic thresholds, or rule-based pipelines that may not generalize across laboratories. In tiling scenarios, duplicate detections in overlap regions and inconsistent fusion of multi-scale outputs can skew counts and downstream summaries. Moreover, per-cell outputs may not provide a standardized payload (coordinates, class taxonomy, confidence) suitable for downstream feature extraction and audit.

[0013] At the feature level, many systems focus on detection probabilities rather than explicitly quantifying nucleus and cytoplasm morphology or chromatin texture in physical units. When segmentation is absent or features are expressed in arbitrary pixels, measurements become scanner-dependent. Spatial organization of abnormal cells (e.g., focal clusters versus diffuse patterns) is also underused in forming a slide-level assessment.

[0014] At the slide level, region-wise scores are sometimes combined by simple averaging. Without probability calibration, such scores can be poorly aligned with true risk, complicating operating-point selection and triage policies. In addition, automatic exclusion of likely negative slides, subject to a bound on false-positive rate, requires calibrated outputs and auditable thresholds, which are not consistently available.

[0015] From a workflow standpoint, common viewers overlay detections on the WSI but may lack interactive galleries of per-cell thumbnails with real-time filtering (by class, confidence or viewport) and mechanisms for capture of expert confirmations or rejections that immediately influence case priority. Where expert feedback is recorded, it is not always persisted in a structured form suitable for systematic retraining and cross-site harmonization.

[0016] Finally, many datasets and validation studies are single-site or limited in platform diversity, making it difficult to establish quality-control bands and operating points that transfer across scanners, laboratories and LBC platforms. There remains a need for methods and systems that address at least some of the following issues: scale harmonization to a common physical unit system, stain appearance standardization and debris handling, multi-level, auditable quality control tied to triage, standardized instance outputs amenable to per-cell morphological and texture analysis, slide-level aggregation that incorporates distributional and spatial descriptors, calibrated probability estimates supporting controlled automatic exclusion policies, and user interfaces that capture expert input in a way that can be used to iteratively improve performance across sites, particularly in the presence of LBC- specific variability.

[0017] Unlike histology systems that reason over tissue architecture, LBC cytology requires instance-level analysis of single cells in physical units, platform-aware quality control tied to triage, and calibrated slide-level probabilities that support automatic exclusion policies. The invention addresses these cytology-specific needs with standardized per-cell outputs, explicit morphology and chromatin-texture features in pm-based units, spatial-distribution descriptors, and a GUI that captures structured accept or reject feedback for immediate triage impact and later retraining.

[0018] SUMMARY OF THE INVENTION

[0019] The present invention addresses the need for an integrated diagnostic pipeline that effectively handles large image datasets, ensures image quality, and provides accurate and reliable diagnoses for cytology microscopic images using whole-slide imaging (WSI). The pipeline couples multi-stage quality control with advanced artificial intelligence (Al) models and explicit feature engineering, yielding a calibrated slide-level probability of dysplasia suitable for triage and reporting.

[0020] In one aspect, each incoming WSI is harmonized to a canonical micrometers-per-pixel resolution to make measurements comparable across scanners and laboratories. Preprocessing further includes stain normalization and debris suppression using a foreground mask, thereby stabilizing appearance and removing non-cellular artefacts before analysis.

[0021] Quality control operates at the slide, tile, and diagnosis levels. Quantitative QC metrics, including brightness, colorfulness, focus / contrast, stain-range conformance, foreground / tissue coverage and cellularity proxies, are evaluated in canonical units and compared with quantile bands learned from a training distribution. Values outside central bands raise WARNING (caution indicator, attenuated confidence, lowered priority); values outside acceptance bands raise ERROR, which blocks automated diagnosis and requests reacquisition. QC bands may be configured per site and per LBC platform. For detection, the pipeline employs a customized YOLO-family neural network with convolutional backbones and multi-scale heads to localize individual cell instances. Alternatively, vision transformers (ViTs) with spatial attention may be used. The detector outputs, for each cell, a payload comprising image coordinates and box dimensions in canonical pixels, a class and a confidence; per-class NMS and multi-scale fusion suppress duplicates. An important aspect related to novelty of the present invention is the subsequent processing and visualization of its outputs.

[0022] Beyond detector confidence scores, the invention derives hand-engineered per-cell features from nucleus / cytoplasm segmentation within a defined neighborhood at the canonical resolution. Features include nuclear area, nucleus-to-cytoplasm (N / C) ratio, circularity and eccentricity, chromatin texture, and local cell density. Computing these quantities in canonical units and on stain-normalized channels reduces site effects and improves interpretability.

[0023] Per-cell attributes are aggregated into a fixed-length slide-level feature vector. For each numeric attribute, the system forms distribution summaries, top-K order statistics and coarse histograms, computed over all cells and per class. To capture tissue organization, spatial- distribution descriptors are added, such as kernel-density percentiles and peak-to-median ratios, DBSCAN cluster count / size / fraction-in-clusters, nearest-neighbor distance statistics, and a spatial autocorrelation score. Where appropriate, features are normalized by foreground area or total cell count.

[0024] The slide-level vector is input to a gradient-boosted decision-tree classifier. The raw score is calibrated on a held-out set to yield a well-calibrated probability of dysplasia. An operating threshold may be selected to maximize automatic exclusion of likely negative slides subject to a user-specified false-positive-rate bound; automatically excluded cases are logged for audit. The QC state (OK / WARNING / ERROR) gates presentation and triage, with ERROR blocking automated output.

[0025] A client user interface renders the WSI alongside a gallery of per-cell thumbnails annotated with classes and confidence scores. The gallery is filterable in real time by class, confidence threshold and current viewport; a top-K control highlights the highest-scoring cells. Per- thumbnail accept / reject actions are captured as structured labels that immediately update the slide-level risk score (probability of dysplasia) and reorder the worklist without reloading the WSI, and are persisted for model improvement.

[0026] The invention further provides an iterative data-expansion and harmonization loop. Detector proposals on additional slides are reviewed by experts to grow a verified multi-lab dataset. All slides are resampled to the canonical MPP and undergo stain normalization; training and validation are site-stratified to ensure robustness across scanners and LBC platforms. QC bands and calibration can be refreshed within validated ranges while maintaining versioning and audit.

[0027] In some embodiments, the pipeline concurrently screens for infectious agents and infection- associated cytopathic changes commonly encountered in PAP cytology. The detector’s class taxonomy can be further extended to include organism- or surrogate-level classes such as Trichomonas vaginalis, Candida (yeast and pseudohyphae), bacterial overgrowth patterns including clue cells / Actinomyces-like organisms, and viral cytopathic effect suggestive of herpes simplex virus (HSV) (multinucleation, molding, chromatin margination). For each detection, the system may emit the standard per-instance payload in canonical units and derive additional features appropriate to the target, such as skeleton-based length / branching for pseudohyphae, aspect ratio and motile-shape descriptors for trichomonads, OD / texture statistics on organismal channels, and co-localized host-cell markers (e.g., clue-cell haloing). These features can be aggregated into infection-specific slide-level summaries and calibrated auxiliary probabilities (e.g., P(Trichomonas), P(Candida), P(BV / Actinomyces), P(HSV-CPE)) that may be reported in a similar manner as P(Dysplasia). LBC-aware preprocessing (stain normalization and debris suppression) and QC bands are tuned to separate organisms from mucus / precipitate and to mitigate focus-spread effects. The GUI may expose per-class filters and a thumbnail gallery for infectious findings, and per-thumbnail accept / reject feedback is persisted to iteratively refine both organism detection and calibration across sites.

[0028] In further embodiments, the method can be stain-agnostic and supports cytology images prepared with stains other than Papanicolaou, including hematoxylin-eosin (HE) and Romanowsky-type stains (e.g., Giemsa / May-Grunwald-Giemsa). During preprocessing, stain normalization may use a per-stain profile: the system may estimate stain vectors for the active stain family and project the WSI into a stain-specific basis (hematoxylin / eosin for HE; azure- eosin-methylene blue for Giemsa), after which detection and feature extraction may operate on harmonized channels. The “hematoxylin-dominant channel” referenced herein can be interpreted generically as a nucleus-enhancing channel within the chosen stain basis. QC bands and thresholds are selected from per-stain configurations to accommodate expected color ranges, contrast and focus spread. The detector’s class taxonomy and calibration can be loaded from stain-specific weights to preserve performance and enable cross-stain comparability; slide-level features remain defined in canonical physical units and therefore unchanged. The GUI may record the stain type from metadata or user input and render stain-appropriate overlays; per-stain accept / reject feedback can be persisted for subsequent retraining and calibration refresh across laboratories and preparations.

[0029] By addressing the challenges of handling large WSI data, ensuring image quality in canonical physical units, combining instance-level Al with interpretable, hand-engineered features and spatial descriptors, and delivering a calibrated, policy-aware slide-level probability with interactive review, the invention significantly enhances the efficiency and reliability of digital pathology diagnostics for cytology microscopic images. The methods may be implemented for automatic diagnosis or as decision support for medical examiners, and may be realized as a system and / or as a non-transitory computer-readable medium storing instructions to execute the steps described.

[0030] In one aspect, the invention provides a computer-implemented method for assisting diagnosis of dysplasia from a liquid-based cytology whole-slide image (WSI), using a multi-stage pipeline combining deep-learning-based detection with engineered feature extraction and a calibrated machine-learning classifier. The method involves receiving and resampling a digital WSI to a canonical spatial resolution, detecting individual cell instances with a neural network, computing morphological and texture features for each detected cell, aggregating these features across the WSI to form a slide-level feature vector, inputting this vector into a calibrated classifier to yield a slide-level probability of dysplasia, computing slide-quality metrics to govern downstream triage, presenting results via an interactive graphical interface with per-cell thumbnails, and iteratively refining the system with expanded and harmonized training data from expert annotations.

[0031] In a preferred embodiment, the morphological and texture features include nuclear area, nucleus-to-cytoplasm ratio, shape descriptors such as circularity and eccentricity, chromatin texture descriptors, and a local cell-density measure. The slide-level feature vector may include statistical summaries such as mean, median, variance, inter-quartile range, specified percentiles, and spatial-distribution descriptors.

[0032] Another preferred aspect involves computing slide-quality metrics, including cell-density, focus, and contrast metrics, assigning a quality state by comparing metrics to predefined quantile bands, and controlling downstream operations based on this state. A WARNING state results in cautionary indicators and lower triage priority, while an ERROR state blocks automated diagnosis and requests reacquisition. An operating threshold may also be selected to exclude likely negative slides automatically, logging such cases for audit.

[0033] In another embodiment, training data is iteratively expanded through manual annotations and automated proposals from the detection network, refined by expert feedback, and harmonized across multiple laboratories by resampling to a common resolution and applying stain normalization. The network and classifier training processes include strategies like focal loss and minority-class oversampling.

[0034] A further preferred embodiment provides a graphical user interface (GUI) displaying a pan- and-zoom view of the WSI along with an interactive gallery of per-cell thumbnails annotated with confidence scores. Selection of thumbnails recenters the slide view, and the gallery supports real-time viewport-based filtering to reduce computational load. The GUI accepts per-thumbnail confirmations or rejections, immediately updating the slide-level probability and worklist order without requiring WSI reload.

[0035] Another embodiment enhances the GUI with class-quota inputs to display top-K thumbnails per class and confidence-threshold controls, where user actions emit structured label events for instantaneous recalculation of risk scores and persistent storage for model retraining.

[0036] Preferably, stain normalization employs a Macenko-type optical-density procedure, complemented by foreground / tissue masking in HSV or OD spaces, with small-object removal to suppress debris and mucus commonly found in cytology preparations.

[0037] An additional embodiment divides WSIs into overlapping, fixed-size tiles prioritized by foreground coverage. This includes generating multi-scale context versions of tiles, allowing the detector to integrate broader contextual information while maintaining computational efficiency.

[0038] Another preferred aspect addresses duplicate detections by applying per-class nonmaximum suppression and confidence-weighted box voting for multi-scale fusion. Slide-level de-duplication across tile overlaps favors instances originating from higher-quality tiles.

[0039] Segmentation in another embodiment occurs within enlarged detection windows, employing seeded watershed methods on hematoxylin-dominant channels or using compact neural networks such as UNet.

[0040] A further preferred embodiment aggregates per-cell features by computing numeric summaries, top-K order statistics, and spatial descriptors, including kernel-density estimates, clustering measures, nearest-neighbor distances, and spatial autocorrelation metrics. Contributions from marginal-quality tiles may be down-weighted for robustness.

[0041] Another embodiment encapsulates these methods into a non-transitory computer-readable storage medium storing instructions executable by processors to implement the diagnostic pipeline.

[0042] Finally, the invention provides a diagnostic system comprising modules for receiving and resampling WSIs, cell detection, feature extraction, feature aggregation, slide-level classification, quality control, interactive GUI presentation, and iterative training-data expansion and harmonization, facilitating reliable dysplasia diagnosis from liquid-based cytology images.

[0043] These and other features, aspects, and advantages of the invention will become better understood with reference to the following drawings, description, and claims.

[0044] NOTATION AND NOMENCLATURE

[0045] For the purposes of this specification, the following terms and conventions apply. Unless stated otherwise, values are exemplary and non-limiting, and measurements are expressed at the canonical spatial resolution.

[0046] Micrometers-per-pixel (MPP): the physical pixel size of the scanner. If MPP metadata are unavailable, MPP may be estimated from scanner profiles, embedded scale bars, or anatomical priors (typical cell sizes).

[0047] Canonical spatial resolution: a target scale to which all WSIs are resampled to enable crossscanner comparability. In one embodiment the canonical resolution is 0.450 pm / px (±0.04) at 20*. All lengths, areas, windows and thresholds below are defined in canonical units.

[0048] Coordinate system: a 2-D rightward-downward image coordinate system with origin at the upper-left of the slide. A stored transform Traw^canonical maps raw scanner coordinates to canonical coordinates and vice versa.

[0049] Tiles: fixed-size, overlap-aware regions extracted from the canonical WSI. By way of example, 1280 x 1280 px tiles correspond to -576 pm x 576 pm at 0.450 pm / px, with -10% overlap expressed in pm at the canonical scale. Each tile records its top-left canonical coordinates, field-of-view in pm, and quality summaries.

[0050] Context versions (multi-scale inputs): for a given tile, additional inputs that preserve the network’s input size while enlarging the field-of-view (e.g., “1 / 4” where the original tile occupies the central quarter; “1 / 16” where it occupies the central sixteenth).

[0051] Foreground / tissue mask: a binary mask indicating tissue / cell-bearing regions, typically derived in HSV or OD space by removing glass background and applying morphological smoothing. Foreground coverage is the fraction of the slide (or tile) area labeled as foreground. Cells-per-area is reported per mm2of foreground.

[0052] Stain normalization: normalization of stain appearance to reduce inter-site variability. In one embodiment a Macenko method is used, with Reinhard or Vahadane as alternatives. “Stain vectors” denote the estimated basis in OD space.

[0053] Liquid-based cytology (LBC): preparations (e.g., ThinPrep, SurePath) characterized by monolayer-like distributions and variable cellularity, debris and stain hue. LBC-specific priors may include debris suppression and minimum colorful-tile quotas; platform-specific QC bands may differ between ThinPrep and SurePath as noted herein.

[0054] Detector output (per-cell payload): for each candidate cell instance the object detector emits (x,y,w,h) in canonical pixels, class 6 {LSIL, HSIL, SCC, AGC-NOS, AGC-NEO, AIS, ADC A, Endocervical, . . . and confidence E [0,1],

[0055] Per-class non-maximum suppression (NMS) is applied at a specified intersection-over-union (e.g., loU 0.5); multi-scale proposals may be fused by confidence-weighted box voting.

[0056] Neighborhood window: a circular neighborhood around a detection used for local statistics; in one embodiment a 50 pm radius at the canonical resolution.

[0057] Segmentation masks: nuclear and cytoplasmic regions obtained by seeded watershed or a compact UNet within the detection window, optionally refined by morphological postprocessing.

[0058] Per-cell morphological features: include nuclear area (pm2), cytoplasmic area (pm2), N / C ratio =AN / AC , circularity =47tA / P2, and eccentricity from second central moments.

[0059] Per-cell chromatin-texture features: grey-level co-occurrence matrix (GLCM) statistics (e.g., contrast, entropy, homogeneity) computed on a hematoxylin-dominant channel over fixed windows in canonical pixels.

[0060] Local cell density: the count of neighboring detected cells with centers within the defined neighborhood radius (e.g., 50 pm).

[0061] Slide-level feature vector: a fixed-length vector produced by aggregating per-cell attributes over all cells and per class, comprising count, mean, standard deviation, median, interquartile range, p5 / p95, min, max, coarse histograms, and top-K order statistics (e.g., mean / min of the highest-scoring K cells). Abnormal-load indicators (e.g., abnormal-to-all ratio, number of abnormal regions) may be included.

[0062] Spatial-distribution descriptors: features derived from cell coordinates, such as kernel-density percentiles and peak-to-median ratios, DBSCAN cluster count / size / fracti on-in-clusters (a and minPts in pm), nearest-neighbor distance statistics, and a spatial autocorrelation score (e.g., Moran’s I). Where relevant, values are normalized by foreground area or total cell count.

[0063] Contrast (tile / slide): global RMS contrast of a luminance channel with dynamic range defined by the lst-99th percentiles; local contrast is the mean of per-tile RMS contrasts; edge-based contrast includes variance of the Laplacian and Tenengrad scores.

[0064] Quality-control (QC) bands and states: for each QC metric the system stores a central quantile band (e.g., 10th-90th percentile) used to raise WARNING, and an acceptability band used to raise ERROR. OK: within central bands; WARNING: outside central but within acceptability (caution indicator, attenuated confidence, lowered priority); ERROR: outside acceptability (automated diagnosis blocked; reacquisition or manual review requested).

[0065] Calibrated probability of dysplasia: P(Dysplasia), the slide-level risk score obtained by inputting the slide-level feature vector to a gradient-boosted decision-tree classifier and applying probability calibration (e.g., isotonic regression) on a held-out set. Detector confidence values are distinct from the calibrated slide-level probability.

[0066] Operating point / automatic exclusion: a decision threshold on P(Dysplasia) selected to maximize the proportion of negative slides automatically excluded from routine review subject to a user-specified false-positive-rate bound; automatically excluded cases are logged for audit.

[0067] BRIEF DESCRIPTION OF DRAWINGS

[0068] The invention is shown by means of example embodiments in a drawing, wherein:

[0069] Fig. 1 illustrates an overview of the diagnostic pipeline;

[0070] Fig. 2 illustrates the division of the WSI into overlapping tiles;

[0071] Fig. 3 illustrates an object-detection neural network for Al detection of tile contents;

[0072] Fig. 4 illustrates a graphical user interface for presenting a final diagnosis;

[0073] Fig. 5 illustrates a functional diagram of the diagnostic analysis system;

[0074] Fig. 6 illustrates an embodiment of a hardware configuration of the system.

[0075] Fig. 7 illustrates receiving a WSI and harmonizing its spatial scale;

[0076] Fig. 8 illustrates per-cell segmentation and feature extraction around detector outputs;

[0077] Fig. 9 illustrates slide-level feature aggregation;

[0078] Fig. 10 illustrates the slide-level classifier and probability calibration pipeline;

[0079] DETAILED DESCRIPTION

[0080] The following detailed description is of the best currently contemplated modes of carrying out the invention. The description is not to be taken in a limiting sense, but is made merely for the purpose of illustrating the general principles of the invention.

[0081] The description presented below relates to an example embodiment of processing Papanicolaou (PAP) smears. However, it is noted that the invention can be used in a similar manner for other types of cytology microscopic images.

[0082] Fig. 1 provides an overview of the diagnostic pipeline for digital pathology using whole-slide imaging (WSI). The pipeline comprises the following steps: Step 110: receiving and harmonizing the WSI. A whole-slide image (WSI) 10 is received from a scanner or database and its spatial scale is harmonized to a canonical micrometers-per- pixel resolution to enable consistent downstream processing across scanners and laboratories.

[0083] Step 110A: stain normalization and debris suppression. The WSI is normalized for stain appearance and foreground is refined by suppressing small debris and mucus, thereby reducing inter-site variability and improving robustness of later analysis.

[0084] Step 111: whole-slide quality control. Global quality metrics (including brightness, colorfulness, focus / contrast and cellularity proxies) are computed in canonical units and compared to learned bands to assign a QC state that governs downstream triage and acceptability.

[0085] Step 112: division into tiles. The harmonized WSI 10 is divided into overlapping tiles 11 at the canonical resolution to make subsequent computations tractable while preserving cellular detail near tile borders.

[0086] Step 113: tile-level quality control. Each tile 11 is evaluated for sufficient information content and visual quality; tiles failing preset criteria are corrected where possible or excluded from further analysis.

[0087] Step 114: object detection of cell instances. An object-detection neural network analyzes each accepted tile to localize individual cell instances and outputs, for each instance, a predicted class, confidence and image coordinates.

[0088] Step 114A: per-cell segmentation and feature extraction. For each detected instance, nucleus and cytoplasm are segmented and a neighborhood at canonical resolution is analyzed to derive morphological and chromatin-texture features and local density measures.

[0089] Step 115: consolidation of per-tile predictions. Per-cell outputs and associated metadata are consolidated across tiles, with de-dupli cation in overlaps and harmonization of coordinates at the slide level.

[0090] Step 115A: slide-level feature aggregation. Per-cell attributes are aggregated into a slidelevel feature vector comprising statistical summaries across all cells and per class, together with spatial-distribution descriptors of the cell pattern on the slide.

[0091] Step 115B: classification and calibration. The slide-level feature vector is input to a gradient-boosted decision-tree classifier and its raw output is calibrated to yield a slide-level probability of dysplasia (P(Dysplasia)).

[0092] Step 116: diagnosis-level quality control. Sanity checks on normal and abnormal cell counts and other diagnostic parameters are performed; the QC state influences confidence display, prioritization and, where applicable, blocking of automated diagnosis. Step 117: final presentation. The calibrated P(Dysplasia) and associated findings are produced for the whole slide and presented to the user in a graphical interface including per-cell thumbnails and interactive review tools.

[0093] Step 118: iterative data expansion and harmonization. Detector proposals on additional slides are accepted or rejected by experts to grow the training set; slides from multiple sites are standardized to the canonical resolution and normalized for stain to sustain cross-lab performance.

[0094] Now, a detailed description of each step will be presented.

[0095] Step 110: receiving and harmonizing the whole-slide image (WSI)

[0096] Step 110 provides the input to the diagnostic pipeline and establishes a common spatial reference for all downstream computations. In this step, a whole-slide image (WSI) 10 is received from an input device, such as a high-resolution slide scanner, or from an image database. The input device captures the entire glass slide at high resolution, producing a digital image that can have a size of up to 200 000 x 200 000 pixels. To ensure consistent processing across scanners, laboratories and acquisitions, the WSI 10 is then harmonized to a canonical micrometers-per- pixel (MPP) resolution before further analysis.

[0097] In one embodiment, the system reads the scanner’s native MPP metadata and computes an affine scale factor, such as s = MPPnative / MPPcanonical. The WSI is resampled to a canonical spatial resolution of approximately 0.450 pm / px (±0.04), corresponding to 20 optical magnification, using bicubic interpolation with anti-alias pre-filtering for down-sampling. If the input is a multi-resolution pyramid, the system selects an appropriate level and applies a precise fractional scale to reach the canonical MPP. The resampling preserves slide orientation and a consistent 2-D coordinate system with origin at the upper-left corner. A transformation record is stored with the slide to enable lossless mapping between raw and canonical coordinates during tiling, per-cell thumbnail generation and user interaction.

[0098] If MPP metadata are missing or unreliable, the system estimates MPP from one or more of: (i) a scanner profile, (ii) verification against embedded scale bars, and (iii) an anatomical- prior estimator that compares observed cell sizes to expected physical ranges in cytology. The selected estimate is validated against internal consistency checks; if the uncertainty exceeds a predefined bound, the case is flagged for quality review as described in later sections. All tile sizes, neighborhood windows and quality-control thresholds referenced below are defined in canonical units, making results comparable across scanners and sites. For clarity, references to pixel sizes in later steps are expressed at 0.450 pm per pixel unless otherwise stated; the stored raw-to-canonical transform is applied whenever coordinates or dimensions cross module boundaries.

[0099] Prior to resampling, the system optionally converts the WSI to a standard color space (e.g., sRGB) and normalizes photometric interpretation so that intensity and color channels are consistent across sources. The output of Step 110 is a harmonized WSI IO C at the canonical MPP together with the stored transformation Traw^canonical and updated metadata 10M. This harmonization reduces downstream computational variability, ensures that tile dimensions correspond to fixed physical extents (e.g., a 1280 x 1280 px tile represents -576 pm x 576 pm at 0.450 pm / px), and enables uniform definition of morphological and texture features. Figure 7 schematically illustrates the receiving and harmonization flow.

[0100] Step 110A: stain normalization and debris suppression

[0101] Step 110A reduces inter-site and inter-platform variability in color appearance and removes non-cellular artefacts that commonly occur in cytology preparations, thereby improving the robustness of later quality control and ensuring that subsequently computed features, particularly chromatin texture, are dependent on cell morphology rather than staining artefacts. In this step, the harmonized WSI IO C undergoes stain normalization and foreground refinement.

[0102] In one embodiment, stain normalization follows a Macenko-type procedure based on the optical-density (OD) transform. The system estimates dominant stain vectors from the slide, projects the image into the stain basis, and reconstructs it to match a stored reference stain distribution; Reinhard color normalization (Lab-domain mean / variance matching) or Vahadane sparse non-negative matrix factorization can be used as fallbacks. The normalization is applied uniformly to the WSI so that chromatin-related intensities and colors used later for texture features are comparable across laboratories and scanners. The selected normalization parameters are stored with the slide to ensure deterministic regeneration of thumbnails and features.

[0103] Foreground refinement proceeds by constructing a slide-level foreground mask in the HSV color space. Background glass is removed by thresholding saturation and value, followed by morphological opening / closing to smooth the mask. Small connected components below a minimum area in physical units (for example, A min in the range of tens of pm2) are suppressed to reduce debris and mucus, which are frequent in cytology preparations. The resulting mask defines the tissue / cell-bearing regions to be tiled and analyzed. In the specific case of liquidbased cytology (LBC), where monolayered distributions with variable cellularity are common, the system can apply LBC-specific priors, including: (i) enforcing a minimum colorful-tile quota to avoid acellular fields, and (ii) adjusting small-object removal and mask-smoothing parameters to the expected particle size distribution of LBC preparations. Platform-specific presets may be used so that ThinPrep and SurePath slides receive appropriately tuned parameters consistent with their typical layer thickness and focus spread.

[0104] The outputs of Step 110A are a stain-normalized image IO N, a refined foreground mask and associated parameter records. These artefacts are consumed by the quality-control procedures (Step 111), by tiling (Step 112), and by the per-cell analysis stages, ensuring that downstream measurements, brightness, colorfulness, chromatin texture and local density, are computed on visually standardized content with non-cellular artefacts minimized.

[0105] Step 111 : Whole-Slide Image Quality Control

[0106] Step 111 in the diagnostic pipeline is dedicated to ensuring the quality of the whole-slide image (WSI) 10 at the canonical spatial resolution established in Step 110, before it undergoes further processing and analysis. This step is important because the accuracy and reliability of the diagnostic results heavily depend on the quality of the input image. Poor-quality images can lead to incorrect diagnoses, which can have serious implications for patient care. Therefore, the quality control process in Step 111 involves a comprehensive evaluation of image-level metrics, computed in canonical units, and assignment of a quality state that governs downstream triage and acceptability of automated diagnosis.

[0107] In this step, metrics are computed on the harmonized image IO C and / or the stain- normalized image IO N as explicitly indicated

[0108] In general, this step determines whether the whole-slide image meets predefined wholeslide quality criteria derived from a training distribution collected across multiple laboratories and scanners. The system evaluates: brightness and colorfulness, focus / contrast, stain-quality ranges, foreground / tissue coverage, and cellularity proxies. Each metric is summarized as a scalar or a small panel of statistics (e.g., mean, interquartile range, tail percentiles) for the entire WSI and, where useful, per region.

[0109] For the purposes of this specification, the following terms are defined as follows (all computed in, or mapped to, the canonical MPP):

[0110] Colorfulness: a measure of the intensity and variety of colors in an image, computed in HSV or an OD-based stain space. Higher values indicate more vibrant and diverse colors.

[0111] Brightness: the overall lightness or darkness of an image, measured on the luminance or HSV value channel. Focus / blur: sharpness indicators computed from edge / gradient responses; higher blur implies reduced sharpness.

[0112] Contrast: separation of intensities within structures; assessed globally and locally.

[0113] Stain quality: coverage and range of stain channels against expected Pap stain distributions. Foreground / tissue coverage: fraction of slide area containing tissue / cells as opposed to glass background.

[0114] Cellularity proxies: measures that correlate with the presence and density of cells.

[0115] The image quality and exposure metrics assessed during this step include colorfulness and its variance, brightness and its variance, and blur and its variance. The system flags images if the colorfulness is too low or too high. Low colorfulness may indicate under-staining or insufficient exposure; high colorfulness may suggest over-staining or dye precipitate that can obscure important details. Brightness is checked against acceptable bands to detect under- or overexposed scans. The sharpness of the WSI is evaluated to detect blurriness that might indicate focus issues or surface impurities.

[0116] Explicit contrast metrics are evaluated at multiple scales. Global contrast is computed as root-mean-square (RMS) contrast on the luminance channel with dynamic range measured between the 1st and 99th intensity percentiles to reduce sensitivity to outliers. Local contrast is the mean of per-tile RMS contrasts and the proportion of tiles whose RMS exceeds a minimum acceptable level. Edge-based contrast is assessed by the variance of the Laplacian and a Tenengrad score obtained from Sobel gradients; these measures respond to hazy or low-contrast scans even when average brightness is nominal.

[0117] Foreground / tissue coverage is obtained from the WSI 10_N (after Step 110A) by thresholding saturation and value in HSV to remove background glass, followed by morphological opening / closing to smooth the mask. The resulting mask is used to compute the percentage of slide area containing tissue. Extremely low coverage indicates sparsely cellular or mis-scanned slides.

[0118] Cellularity metrics are estimated using two complementary proxies available at slide-level QC time: (i) the fraction of “colorful tiles” (tiles whose saturation and local variance exceed minimum thresholds) which correlates with cellularity; and (ii) cells-per-area estimates when a lightweight blob / nuclei proxy or fast detector pass is enabled. These metrics are summarized globally and compared to the training distribution to detect under-cellular or overloaded preparations typical of inadequate liquid-based cytology. Stain-quality metrics are assessed using HSV and / or OD stain channels. The system verifies that stain vectors and per-channel intensities fall within expected Pap ranges; substantial deviations can indicate staining or scanner color management failures.

[0119] All QC metrics are compared to quantile bands learned from a curated multi -lab training set. For each metric the system stores: (i) a central quantile band (e.g., 10th-90th percentile) that, when exceeded, triggers a WARNING; and (ii) a diagnostic-acceptability band that, when breached, triggers an ERROR. Multiple findings can co-exist; the highest severity governs downstream behavior. In particular:

[0120] OK: all metrics within central bands; Al diagnosis proceeds normally.

[0121] WARNING: one or more metrics outside central bands but within acceptability; a yellow caution indicator is surfaced, confidence displays may be attenuated, and the case’s triage priority is reduced, but Al inference is not blocked.

[0122] ERROR: one or more metrics outside acceptability bands (e.g., focus or cellularity below validated minimums, or stain hue far outside Pap ranges); automated diagnosis is blocked and re-acquisition or manual review is requested.

[0123] Where site effects matter (for example, ThinPrep vs SurePath focus spread), the same policy is applied with platform-specific bands: tighter density and focus bands on ThinPrep; a slightly wider focus band on SurePath to accommodate layer-thickness variability. The system may perform controlled updates of the QC thresholds, particularly the WARNING bands, based on ongoing analysis of historical data. Any such adjustment is version-controlled and logged to preserve full auditability and ensure diagnostic consistency over time.

[0124] The quality control process begins with automated measurement of the above metrics, proceeds to state assignment (OK / WARNING / ERROR), and records a QC summary object that includes metric values, bands used, state, and recommended actions. Depending on the issue, the system might apply corrective image processing (e.g., brightness or color balance adjustment) or recommend re-scanning. Images that cannot be corrected to meet the quality standards are excluded from further automated analysis. The QC state is propagated to later stages and influences triage, confidence presentation and blocking behavior as described in subsequent steps. All QC outcomes, including any automatic exclusions, are logged for audit.

[0125] By implementing a rigorous quality control process in Step 111, the diagnostic pipeline ensures that only high-quality WSIs are used for further analysis. This step is important for achieving accurate and reliable diagnostic results, ultimately improving patient care and outcomes. Step 112: Division of the Scan into Smaller Regions

[0126] Fig. 2 illustrates the initial tiling step of the diagnostic pipeline, where the harmonized whole-slide image (WSI) 10 is divided into smaller, manageable regions to facilitate Al prediction. This step is important for handling the immense resolution and data size of wholeslide images, ensuring that the subsequent Al analysis is both efficient and accurate.

[0127] The entire whole-slide image of a PAP smear, typically very large, encompasses a high- resolution scan of the entire slide. After harmonization to the canonical micrometers-per-pixel resolution, the WSI 10 is divided into smaller, equally sized regions called tiles 11 whose dimensions are defined in canonical units. For instance, each tile might be 1280 x 1280 pixels, which corresponds to approximately 576 pm x 576 pm at 0.450 pm / px. The tiles overlap each other by 10% (e.g., 128 pixels, expressed and enforced in pm at the canonical scale). This overlap is large enough to capture cells close to tile edges while limiting duplication. If the same cell is encountered in overlap regions, it is subsequently removed using non-maximum suppression algorithms during consolidation.

[0128] Tiling is aligned to a fixed slide-level coordinate system (origin at the upper-left comer). Each tile is assigned a unique identifier and metadata 10M including its top-left canonical coordinates, field-of-view size in pm, foreground coverage fraction, and coarse visual summaries (mean brightness, colorfulness, optional focus proxy). The foreground mask derived in Step 110A is used to skip acellular regions proactively, reducing computational overhead. In one embodiment, the tiling scheduler prioritizes tiles with higher foreground coverage or stronger informational content to accelerate early, informative detections.

[0129] On these tiles 11, individual cells 12 can be identified in the subsequent steps of the process. To support multi-scale context during detection, for each tile (imaged at the highest magnification, 1 / 1) the system also prepares context versions representing larger fields of view rendered at the network’s input size. In one embodiment, a 1 / 4-scale context is produced such that the original tile content occupies the central quarter of the input while the surrounding area provides 2x larger x and y extents; similarly, a 1 / 16-scale context provides 4x larger x and y extents with the original tile occupying the central sixteenth. These context views allow the detector to leverage both fine cellular detail and broader morphological cues related to preparation artefacts and staining patterns.

[0130] Depending on slide size, the whole-slide image 10 can be divided into approximately 2,000 tiles using the above parameters, though the exact number varies with WSI dimensions and foreground coverage. The resulting tile set forms a worklist for further processing. Tile coordinates and indices are recorded so that detections can be mapped back to slide-level coordinates for later consolidation, visualization and user interaction. The output of Step 112 is a canonical, overlap-aware grid of tiles 11 (with associated context versions and metadata) ready for tile quality control (Step 113) and Al detection (Step 114).

[0131] By expressing tile geometry and overlap in canonical physical units, the pipeline maintains cross-scanner comparability, ensures that per-cell measurements derived later are physically meaningful, and enables deterministic de-duplication of instances detected across tile boundaries.

[0132] Step 113: Tile Quality Control

[0133] Step 113 in the diagnostic pipeline ensures the quality of each tile 11 after the whole-slide image (WSI) 10 has been divided into smaller regions. This step is important because the accuracy and reliability of the diagnostic results depend on the quality of the individual tiles that make up the entire slide. Poor-quality tiles can lead to incorrect detections and biased slide-level features, which may have serious implications for patient care. Therefore, the tile quality control process in Step 113 involves a comprehensive evaluation of visual and structural metrics computed at the canonical spatial resolution established in Step 110, with acceptance decisions expressed as OK / WARNING / ERROR states that govern downstream processing.

[0134] In general, this step determines whether each tile meets predefined tile-quality criteria derived from a multi-site training distribution. The system evaluates, for each tile 11 : brightness and colorfulness, focus / blur and contrast, stain-range conformance, and foreground (tissue) coverage / cellularity proxies. Metrics are recorded together with the tile’s canonical coordinates to enable audit and reproducibility.

[0135] The tile quality control comprises an assessment of the colorfulness of each tile and its local variance, which is essential for assessing staining sufficiency. Tiles with colorfulness below a minimum threshold may indicate under-staining or acellular fields; excessively high colorfulness may indicate over-staining or precipitate. The brightness of each tile is checked against acceptable bands to detect under- or over-exposed regions that can affect diagnostic accuracy. The sharpness of each tile is evaluated to detect blurriness that might indicate local focus issues or surface impurities. In addition to blur measures, contrast is explicitly assessed at the tile level using RMS contrast of the luminance channel and edge-based measures such as variance of the Laplacian and a Tenengrad score computed from Sobel gradients; these respond to hazy or low-contrast tiles even when average brightness is nominal. Stain-quality conformance is evaluated by comparing per-tile stain vectors or channel ranges (in HSV or OD space) with expected Pap stain distributions. Significant deviations are flagged, as they can degrade chromatin-texture features extracted later.

[0136] Foreground coverage is computed by intersecting the tile footprint with the slide-level foreground mask produced in Step 110A. Tiles with negligible foreground (predominantly glass background) are marked as uninformative. As a complementary cellularity proxy, the system tests whether the tile qualifies as “colorful” (minimum saturation and local variance thresholds). In one embodiment, tiles that fail both the foreground and colorful tests are rejected early to reduce computational overhead.

[0137] Where applicable, platform-specific adaptations may be enabled for liquid-based cytology (LBC). For example, parameters for small-object suppression and mask smoothing are tuned to typical LBC debris sizes; cellularity and focus bands may be tighter for ThinPrep and slightly wider for SurePath to account for layer-thickness variability. These adaptations are implemented without changing the canonical -unit definitions, preserving cross-scanner comparability.

[0138] All tile QC metrics are compared to quantile bands learned from the training set and stored as configuration per site and preparation type. For each metric, the engine maintains (i) a central quantile band (e.g., 10th-90th percentile) used to raise a WARNING when exceeded, and (ii) an acceptance band used to raise an ERROR when breached. Multiple findings can co-exist; the highest severity governs the tile’s state:

[0139] OK (accept): All metrics within central bands; the tile proceeds to Al detection (Step 114) with nominal weight.

[0140] WARNING (accept with caution): One or more metrics outside central bands but within acceptance; the tile proceeds to detection, is flagged in metadata, and may receive reduced influence in later aggregation (e.g., down-weighting in feature summaries).

[0141] ERROR (reject): One or more metrics outside acceptance bands (e.g., severe defocus, nearzero foreground, extreme stain shift); the tile is excluded from Al detection. If rejection yield exceeds a configurable proportion for the slide, a slide-level QC escalation is issued as described in Step 111.

[0142] Where feasible, the system attempts automatic remediation before final rejection. Depending on the issue, the engine may apply brightness / gamma adjustment or white-balance normalization and re-evaluate the tile. If the remediated tile meets acceptance criteria, it is admitted with a WARNING flag and the remediation parameters are stored; otherwise the original ERROR stands. All QC outcomes are logged with metric values, thresholds used, and actions taken.

[0143] To support deterministic handling of overlap regions (introduced in Step 112), QC metadata are attached to each tile identifier together with its top-left canonical coordinates, field- of-view size, and summaries (mean brightness, colorfulness, RMS contrast, Laplacian variance, Tenengrad, foreground fraction). These data are used later to (i) prioritize tiles with higher informational content for early inference, (ii) guide de-duplication in overlaps by preferring detections from higher-quality tiles (see step 115), and (iii) annotate the user interface with per- region quality cues.

[0144] By implementing a rigorous tile-level quality control process in Step 113, grounded in canonical units, quantile-band thresholds and platform-aware adaptations, the pipeline ensures that only high-quality, informative tiles are passed to Al detection (Step 114). This step is fundamental to achieving accurate and reliable diagnostic results at the slide level, while enabling traceability and consistent triage behavior across scanners and laboratories.

[0145] Step 114: Al Detection of Tile Contents

[0146] Step 114 involves applying an object-detection neural network 300 to rapidly identify and localize cells that may exhibit cancerous characteristics within each accepted tile 11, as shown in Fig. 3. The detector operates at the canonical spatial resolution established in Step 110 and consumes the stain-normalized content and foreground mask produced by Step 110A. Each tile is analyzed individually to distribute computational load and to preserve deterministic mapping between detections and slide coordinates.

[0147] In one embodiment, the detector is a customized YOLO-family convolutional neural network with multi-scale feature aggregation (e.g., FPN / PAFPN) and heads operating on inputs that combine the tile at its highest magnification (denoted 1 / 1) with context versions rendered as described in Step 112 (e.g., 1 / 4 and 1 / 16 context where the original tile occupies the central portion). This multi-resolution approach allows the network to leverage fine cellular detail and broader contextual cues such as preparation artefacts and staining patterns while controlling computation.

[0148] The network processes each tile through the following functional blocks:

[0149] Initial feature extraction layer 301. Convolutional blocks extract a high-dimensional feature map representing edges, textures and shapes in canonical pixels. Enhanced feature map layer 302. Attention or spatial emphasis mechanisms weight regions of interest, producing an enhanced feature map in which diagnostically relevant structures are accentuated.

[0150] Detection and localization layers 303. Multi-scale heads generate class-specific bounding- box proposals and confidence scores.

[0151] For each candidate cell instance the detector can emit a per-cell payload in canonical coordinates, such as:

[0152] (x,y,w,h), classe{LSIL, HSIL, SCC, AGC-NOS, AGC-NEO, AIS, ADCA, Endocervical, ... }, confidences [0,1] wherein (x,y) denotes the box center (or top-left, per stored convention) and (w,h) the box dimensions, all expressed in canonical pixels. Classes include dysplastic categories and non- dysplastic contextual classes (e.g., Endocervical) to improve downstream specificity.

[0153] To suppress duplicates and harmonize proposals, the system applies per-class nonmaximum suppression (NMS) at loU 0.5 within each scale, followed by multi-scale fusion by confidence-weighted box voting, yielding a single consolidated set of instances per tile. Detections arising in overlap regions between adjacent tiles are later de-duplicated at slide level (Step 115) by NMS using slide coordinates, with a preference for instances originating from higher-quality tiles (Step 113).

[0154] Training data for the detector originate from manually labelled slides and are expanded iteratively by a “machine-teaching” process in which early models propose candidate cells for expert accept / refine / reject decisions (see Step 118). The detector can be trained with classimbalance handling (e.g., focal-style loss or class weights) and augmentations that preserve morphology while varying stain / blur within validated limits. Site-stratified validation is used to ensure robustness across scanners and preparations, including liquid-based cytology (LBC).

[0155] The detector’s per-cell outputs are stored as metadata 13 alongside provenance (tile ID, scale head), QC flags inherited from the source tile, and thumbnails for later presentation. These outputs serve two immediate purposes: (i) they drive the thumbnail gallery and inspection tools in the user interface (see Step 117), and (ii) they seed per-cell segmentation and feature extraction in Step 114A.

[0156] Step 114A: Per-Cell Segmentation and Feature Extraction

[0157] Step 114A derives quantitative descriptors from each detected instance to support slidelevel analysis and classification. Operating at the canonical spatial resolution and on the stain- normalized image, the system segments nuclear and cytoplasmic compartments and extracts morphological and chromatin-texture features, together with a neighborhood-based density measure.

[0158] Segmentation: each detection’s bounding box defines a processing window enlarged by a small margin (e.g., 10-20% of box size in canonical pixels). Within this window the system computes a hematoxylin-dominant channel (OD space or color-deconvolution). In one embodiment, segmentation proceeds by seeded watershed initialized from the detector’s central region and refined by gradient magnitude and morphological reconstruction; in other embodiments, a compact UNet produces nucleus and cytoplasm masks directly. Post-processing enforces shape-plausibility constraints (e.g., minimum / maximum area in pm2, hole filling, smallobject removal) and rejects masks with extreme eccentricity or fragmentation.

[0159] Neighborhood window: to capture local context, a circular neighborhood of 50 pm radius (expressed in canonical pixels) is defined around the instance center. This window is used for texture statistics and local cell-density estimation.

[0160] Per-cell features: from the masks and neighborhood, the following features are computed with units tied to canonical scale:

[0161] Nuclear area AN in pm2.

[0162] Cytoplasmic area AC in pm2; N / C ratio =AN / AC

[0163] Circularity =47tAN / P2, where P is nuclear perimeter.

[0164] Eccentricity from second central moments of the nuclear mask.

[0165] Chromatin texture on the hematoxylin-dominant channel using GLCM statistics (e.g., contrast, entropy, homogeneity) computed over fixed window sizes and offsets in canonical pixels.

[0166] Local cell density: count of neighboring detected cells whose centers fall within 50 pm.

[0167] Intensity-based measures are computed on stain-normalized values to reduce site effects. Features failing plausibility bounds (e.g., areas outside validated ranges) are flagged for caution and can be down-weighted in later aggregation.

[0168] Coordinate and provenance handling: all per-cell features retain: (i) the slide-level coordinates obtained by adding the tile origin to the instance location, (ii) the class and confidence from Step 114, (iii) the source tile QC state, and (iv) references to per-cell thumbnails. This enables deterministic back-tracking in the user interface and quality-aware aggregation.

[0169] The per-cell feature records produced in Step 114A serve as inputs to consolidation (Step 115) and to slide-level feature aggregation (Step 115A), where statistical summaries across all cells and per class, as well as spatial-distribution descriptors, are formed. By computing segmentation and features in canonical physical units and on standardized stain channels, the pipeline ensures that downstream classification and calibration operate on comparable, siteagnostic measurements.

[0170] Step 115: Combination and Consolidation of Predictions

[0171] Step 115 aggregates the outputs produced on individual tiles into a coherent, slide-level set of cell instances and preliminary summaries. This step resolves duplicates arising from tile overlaps, harmonizes coordinates into a single slide reference frame at the canonical spatial resolution, and prepares quality-aware metadata for downstream feature formation and classification.

[0172] Each accepted tile 11 supplies a list of per-cell payloads from Step 114, bounding box (x,y,w,h), class, and confidence, together with provenance (tile identifier, scale head) and the tile’s QC state from Step 113. Using the stored tile origins and the transformation from Step 110, all detections are mapped to slide-level canonical coordinates. Detections that lie within overlap regions between adjacent tiles are de-duplicated by slide-level non -maximum suppression (per class) at a preset loU (e.g., 0.5). When two candidates compete, the engine breaks ties by confidence and, if needed, by tile quality (preferring instances from tiles with higher sharpness / contrast / foreground coverage). Multi-scale proposals that refer to the same cell are already fused within Step 114; Step 115 therefore only reconciles inter-tile duplication.

[0173] The consolidated instance set is stored together with tile / QC lineage and per-instance thumbnails cropped from the canonical WSI. In addition, Step 115 computes lightweight preliminary summaries (e.g., counts per class, mean confidence per class) that are useful for progress reporting and UI previews. Where required for rapid triage, a simple preliminary score may be computed as a weighted combination of instance confidence scores (weights may reflect tile quality and class priors); however, in this embodiment the authoritative slide-level decision is produced by the classifier of Step 115B.

[0174] Finally, Step 115 ensures that each retained instance is linked to its per-cell features generated in Step 114A by a stable identifier. This linkage enables deterministic back-tracking from slide-level statistics to raw image evidence in the graphical interface and supports quality- aware aggregation in the next step.

[0175] Step 115A: Slide-Level Feature Aggregation

[0176] Step 115A converts the set of per-cell detections and features into a fixed-length slidelevel feature vector expressed in canonical physical units. This vector captures both distributional statistics of cellular morphology / texture and spatial-distribution descriptors of the cell pattern over the slide, thereby supplying discriminative inputs for the slide-level classifier.

[0177] Aggregation scope: Aggregation is performed twice: (i) over all detected cells and (ii) per cytology class (e.g., LSIL, HSIL, SCC, AGC-NOS, AGC-NEO, AIS, ADCA, Endocervical). Before aggregation, an optional low-confidence floor may suppress instances with confidence below a configurable threshold, and optional class smoothing may reweight borderline instances near the decision boundary.

[0178] For each numeric per-cell attribute we compute count, mean, standard deviation, median, interquartile range, p5, p95, minimum, maximum, plus coarse histograms and top-K order statistics such as the mean and minimum of the highest-scoring K cells for K in {5, 10, 50}. Spatial descriptors include KDE density percentiles and peak-to-median ratios, DBSCAN cluster count and size distribution with epsilon defined in pm and site-validated minPts, nearest-neighbor distance statistics, and Moran's I. Features are optionally normalized by foreground area or total cell count, and absent classes are imputed with neutral values accompanied by presence flags.

[0179] Distributional statistics: for every numeric per-cell attribute (e.g., nuclear area, N / C ratio, circularity, eccentricity, GLCM contrast / entropy / homogeneity, local neighbor count), the featurizer computes a fixed panel of summaries: count, mean, standard deviation, median, interquartile range (IQR), p5 / p95, minimum, maximum; coarse histograms (e.g., fixed bins for N / C ratio and confidence); top-K order statistics such as the mean and minimum of the highest-scoring cells for Ke{5, 10,50}.

[0180] Abnormal-load indicators are also formed, including total abnormal-cell count, abnormal- to-all ratio, and the number of distinct abnormal regions (derived from clustering, below). Spatial-distribution descriptors: using the slide-level coordinates of detections, the system forms descriptors that distinguish focal hotspots from diffuse patterns: a kernel-density estimate (KDE) over the slide with density percentiles and peak-to-median ratios (bandwidth specified in pm and chosen relative to typical cell spacing);

[0181] DBSCAN clustering in canonical units (a in pm; minPts set empirically) yielding cluster count, average and maximum cluster size, and fraction of cells in clusters; nearest-neighbor distance statistics (mean, median, p5 / p95); a spatial autocorrelation score such as Moran’s I on the point pattern of abnormal cells.

[0182] (Other embodiments may employ Ripley’s K; Moran’s I is used here for compactness and robustness.) Normalization, missing data and robustness: where relevant, features are normalized by foreground tissue area (mm2) or total cell count so that values are comparable across scanners, laboratories and preparations. Missing entries (e.g., when a class is absent) are handled by imputation with neutral values and accompanied by presence flags. Prior to classification, features may undergo robust scaling (e.g., median / IQR) learned on the training distribution.

[0183] The output of Step 115 A is a rich, fixed-length slide-level feature vector. This vector creates a comprehensive representation that captures not only the statistical distribution of cellular attributes but also their spatial organization on the slide, making it a highly discriminative input for robust classifiers like gradient-boosted decision trees. This vector is then supplied to the classifier in Step 115B.

[0184] Step 115B: Classification and Calibration of Slide-Level Risk Score

[0185] Step 115B transforms the aggregated slide-level feature vector into a calibrated Probability of dysplasia P(Dysplasia) and provides thresholds for triage and optional automatic exclusion of likely negative slides, consistent with the quality-control policy.

[0186] Model family and training: In a preferred embodiment, the classifier is a gradient-boosted decision-tree model (e.g., XGBoost, LightGBM, or CatBoost), as this model family is particularly effective at handling the heterogeneous and tabular nature of the slide-level feature vector. Other machine learning classifiers capable of operating on such vectors, such as Random Forests or Support Vector Machines, may also be used. Training labels are derived from reference diagnoses (e.g., consensus cytology or biopsy). Cross-validation is site-stratified to reflect the multi-lab nature of the data and to guard against site leakage. Class imbalance is addressed by cost-sensitive learning and / or sampling strategies; augmentations used upstream preserve morphology while varying stain / blur within validated limits.

[0187] Calibration: raw classifier scores are converted to a calibrated probability via isotonic regression fit on a held-out validation set. Calibration quality may be visualized as a calibration curve (see Fig. 10 in one embodiment). The purpose of this calibration step is to ensure that the output P(Dysplasia) is a reliable and probabilistically meaningful risk score, such that a score of 0.8 corresponds to an 80% likelihood of dysplasia. This reliability is essential for implementing auditable triage policies, such as the operating point selection for automatic exclusion. Calibration quality is verified by plotting predicted probabilities against observed frequencies on a held-out dataset to generate a calibration curve. The calibrated output is reported as the slidelevel P(Dysplasia) used by the user interface and downstream policies. Operating point selection and triage: to support efficient workflows, an operating threshold can be selected to maximize the proportion of negative slides automatically excluded from routine review, subject to a user-specified false-positive-rate bound. Slides below the “autoexclude” threshold are moved to a low-priority queue and logged for audit; slides above the threshold are prioritized by descending P(Dysplasia). The classifier output is combined with the QC state from Steps 111-113:

[0188] OK QC: standard display and triage.

[0189] WARNING QC: confidence presentation may be attenuated and priority slightly reduced while still allowing Al output.

[0190] ERROR QC: automated diagnosis is blocked irrespective of P(Dysplasia); reacquisition or manual review is requested.

[0191] The calibrated probability and triage decision are exposed to the graphical user interface. As described later (Step 117), per-thumbnail accept / reject actions by the user emit structured label events that immediately update the slide-level risk score and reorder the worklist without reloading the WSI; these interactions are persisted and contribute to iterative model improvement (Step 118).

[0192] By consolidating detections (Step 115), summarizing morphology and spatial organization into a robust slide-level vector (Step 115A), and producing a calibrated, policy- aware risk (Step 115B), the pipeline delivers a reliable and actionable diagnostic signal that is consistent across scanners, laboratories and cytology preparations.

[0193] Step 116 in the diagnostic pipeline is dedicated to ensuring the quality and reliability of the comprehensive diagnosis before it is presented. This step is important because it acts as a final checkpoint to verify that all preceding analyses and predictions have been accurate and consistent. Operating at the canonical spatial resolution and consuming the consolidated instances and slide-level features produced in Steps 115 and 115A together with the calibrated probability from Step 115B, the diagnosis quality control process evaluates a panel of metrics and assigns a diagnosis QC state that governs presentation, triage and, where applicable, blocking of automated output.

[0194] In general, this step determines whether the comprehensive diagnosis meets predefined diagnosis-quality criteria derived from a multi-lab training distribution. The system evaluates at least the following:

[0195] Normal-cell count: the total number of detected normal cells is compared against an expected range learned from the training corpus (expressed per mm2of foreground where appropriate). Counts well below the range can indicate under-detection due to poor image quality or insufficient staining; unusually high counts may reflect over-segmentation or a highly concentrated sample.

[0196] Abnormal-cell count: the number of detected potentially dysplastic cells is compared against acceptability bands (for example, an empirical range such as 25-50,000 cells at the slide level, adjusted per preparation and lab). Too few may suggest under-detection; too many may indicate false-positive proliferation or overloaded preparations.

[0197] Distribution consistency: the spatial distribution of abnormal detections (cluster count / size, nearest-neighbor statistics, Moran’s I) is checked for anomalies relative to the case mix. Extreme clustering in very low-quality regions or excessive scatter inconsistent with the image content prompts review.

[0198] Feature coherence: internal consistency checks compare key per-cell features (e.g., N / C ratio, nuclear area, chromatin texture) against the class taxonomy produced by the detector; systematic mismatches (for example, HSIL labels with uniformly low N / C ratios) are flagged. Aggregation sanity checks: top-K order statistics are contrasted with global summaries; large divergences can signal a few spurious outliers dominating the risk estimate.

[0199] QC inheritance: the proportions of detections and high-influence features originating from WARNING / ERROR tiles are examined; results dominated by marginal-quality tiles may be attenuated or escalated.

[0200] All metrics above are compared to quantile bands learned from a site-stratified training distribution. Consistent with prior QC stages, Step 116 assigns one of three states:

[0201] OK: all diagnosis-level metrics within central bands; the calibrated Probability of dysplasia P(Dysplasia) is displayed and used for triage without modification.

[0202] WARNING: one or more metrics outside central bands but inside acceptance; a caution indicator is surfaced, confidence displays may be attenuated, and the worklist priority is reduced, but automated diagnosis is not blocked.

[0203] ERROR: one or more metrics outside acceptance (e.g., abnormal-cell count implausibly low given visible hotspots, or predominant reliance on tiles failing acceptance at earlier QC); automated diagnosis is blocked and reacquisition or manual review is requested irrespective of P(Dysplasia).

[0204] Where feasible, the engine attempts corrective actions prior to final state assignment, for example, re-running per-cell segmentation with more conservative parameters in regions suspected of over-segmentation, or de-weighting contributions from marginal tiles and recomputing summaries. If corrections bring all metrics within acceptance, the state is downgraded to WARNING with remediation details recorded. Otherwise, ERROR is maintained. The full QC outcome (metric values, bands, remediation attempts, final state) is logged for audit and exposed to the user interface as a concise, human-readable panel.

[0205] By implementing a rigorous diagnosis-level quality control process in Step 116, the pipeline ensures that the comprehensive diagnosis is both accurate and reliable. This step is fundamental to achieving high-quality diagnostic results, ultimately improving patient care and outcomes.

[0206] Step 117: Final Presentation

[0207] Step 117 is the culmination of the diagnostic process, where the combined predictions from individual tiles and the slide-level feature vector are synthesized into a final, calibrated Probability of dysplasia P(Dysplasia) and presented to the user. This step provides the ultimate diagnostic output to inform clinical decisions and patient care.

[0208] The system consumes the calibrated classifier output from Step 115B, the diagnosis QC state from Step 116, and the consolidated instance set and per-cell features from Steps 115 and 114A. The final result includes: the slide-level P(Dysplasia) and (optionally) a risk band derived from an operating threshold chosen to maximize automatic exclusion of negative slides subject to a user-specified false- positive-rate bound; counts and proportions per cytology class, top-K abnormal cells, and summary plots (e.g., probability histogram, cluster counts); a concise QC banner indicating OK / WARNING / ERROR and any explanatory messages.

[0209] The results are presented through a graphical user interface (GUI) as illustrated in Fig. 4. The GUI is structured to enable efficient inspection and interactive refinement. A first pane renders a pan-and-zoom view of the WSI at the canonical resolution, with overlays for detection boxes and cluster heatmaps. Selecting an item recenters the view. A second pane displays a gallery of per-cell thumbnails annotated with their class and confidence. The gallery supports real-time filters, including a crucial viewport filter that displays only thumbnails of cells located within the user's current view of the WSI. This provides the significant technical effect of drastically reducing the computational and network load on the client device, as well as cognitive load on the user enabling a fluid user experience even with thousands of detections on a slide. A class-quota input displays per-class top-K thumbnails. Per-thumbnail accept / reject actions are provided. Each action emits a structured label event that is persisted and immediately triggers an incremental update of the slide-level risk score and worklist triage order without reloading the WSI. Accepted / rejected instances are visually flagged in both panes to maintain context. When the diagnosis QC state is WARNING, the GUI attenuates per-cell confidence displays and surfaces remediation hints; when ERROR, automated results are visually blocked and reacquisition guidance is shown while still allowing navigation for manual review.

[0210] The system generates a final report containing the calibrated P(Dysplasia), a summary of detected abnormalities with locations and confidence scores, the diagnosis QC state and any automatic exclusion / triage decisions. All interactions (filters, accept / reject actions) are logged and become available for subsequent model retraining as described in Step 118.

[0211] Step 118: Iterative Data Expansion and Harmonization

[0212] Step 118 describes the continuous-learning loop by which the system improves detector and classifier performance and maintains cross-site consistency. This step is important for sustaining accuracy across laboratories, scanners and preparation methods, including liquidbased cytology (LBC).

[0213] Seed annotation and proposal loop: training begins with manual annotation of at least N cells per dysplasia class on initial slides. A trained detector is then used to propose cell annotations on additional slides. Domain experts review proposals and can accept, refine, or reject them. Accepted annotations are added as positive examples, while explicitly rejected false positives are added as 'hard negative' examples. This process of hard negative mining is crucial for reducing the detector's false-positive rate over time. This loop is repeated to accumulate a verified multi-lab dataset.

[0214] Cross-site harmonization: to reduce inter-site variability, all slides are resampled to the canonical MPP (e.g., 0.450 pm / px ±0.04 at 20*) as described in Step 110 and undergo stain normalization (Macenko, with optional Reinhard / Vahadane) as described in Step 110A. Augmentations used for training preserve cellular morphology while varying stain, blur and brightness within validated limits.

[0215] Model training and validation: the detector is trained with class-imbalance handling (e.g., focal-style loss or class weights). The slide-level classifier (e.g., CatBoost) is trained on the aggregated features from Step 115A with site-stratified cross-validation. Raw scores are calibrated by isotonic regression on a held-out validation set to yield the calibrated P(Dysplasia).

[0216] Deriving QC bands and operating points: the multi-lab corpus is used to estimate training distributions for QC metrics (focus / contrast, stain ranges, cellularity proxies), from which central quantile bands (for WARNING) and acceptability bands (for ERROR) are set. Preparationspecific bands are maintained, for example tighter density and focus limits for ThinPrep and a slightly wider focus band for SurePath to accommodate layer-thickness variability. Separately, an operating point for automatic exclusion of likely negative slides is selected to meet a configurable false-positive-rate bound and is documented for audit.

[0217] Validation example

[0218] In one validation, 12 097 LBC slides from 10 laboratories (10 153 ThinPrep, 1 944 SurePath) were used for development and site-stratified evaluation, yielding a whole-slide classification AUC of approximately 0.93. Platform-aware QC bands were tuned accordingly. These numbers are exemplary and do not limit the scope of the invention.

[0219] Governance, audit and deployment: all trained models, calibration mappings and QC thresholds are versioned. Changes to thresholds or operating points are recorded with effective dates and validation summaries. During deployment, the system can dynamically adjust nonblocking thresholds within validated ranges based on ongoing analysis and historical data (feedback loop) while preserving traceability of each decision. User interactions captured in Step 117 (accept / reject events) are periodically incorporated to refresh the detector and classifier, after which calibration and QC bands are re-verified.

[0220] By iteratively expanding the training set, harmonizing images across laboratories and preparation methods, and maintaining calibrated models with audited QC policies, Step 118 sustains robust performance of the diagnostic pipeline and supports reliable, site-agnostic operation over time.

[0221] Liquid-Based Cytology (LBC) Adjustments

[0222] Liquid-based cytology (LBC) slides (e.g., ThinPrep, SurePath) exhibit relatively monolayered cell distributions with higher between-slide variance in cellularity, debris / mucus and stain hue than conventional smears. The pipeline accommodates these traits by applying LBC-specific priors in preprocessing, quality control and operating-point selection while keeping all measurements in canonical physical units.

[0223] Preprocessing (Steps 110 / 110A).

[0224] All LBC WSIs are resampled to the canonical micrometers-per-pixel resolution to harmonize scale across scanners. Stain appearance is normalized (e.g., Macenko, with Reinhard / Vahadane alternatives), and the foreground / tissue mask is refined in HSV / OD space with small-object removal to suppress debris / mucus typical of LBC. To avoid acellular fields, the engine enforces a minimum colorful-tile quota (tiles exceeding saturation and local -variance minima); tiles failing both foreground and colorful tests are deprioritized or skipped. Quality control bands (Step 111 and tile QC in Step 113).

[0225] QC metrics (brightness, focus / contrast, stain-range conformance, cellularity proxies and cells-per-area) are compared to quantile bands derived from LBC training data. Platform-aware adaptations are applied:

[0226] ThinPrep: tighter density and focus bands reflecting its typical thin monolayer and sharper focus spread.

[0227] SurePath: a slightly wider focus band to accommodate thicker / variable layers.

[0228] Violations of central bands raise WARNING (caution indicator, attenuated confidence, lowered priority); breaches of acceptance bands raise ERROR (automated diagnosis blocked, reacquisition / manual review requested). Bands are stored per site / preparation and may be tuned within validated ranges.

[0229] Detection and features (Steps 114 / 114A).

[0230] The detector processes tiles at canonical resolution and emits per-cell payloads (x,y,w,h), class (e g., LSIL, HSIL, SCC, AGC-NOS, AGC-NEO, AIS, ADCA, Endocervical, ...) and confidence. LBC artefacts are mitigated by the stain-normalized inputs and debris suppression; overlap and multi-scale duplicates are resolved by per-class NMS and box voting. Per-cell nucleus / cytoplasm segmentation and morphological and texture features (nuclear area, N / C ratio, circularity, eccentricity, GLCM texture on the hematoxylin-dominant channel, local density within 50 pm) are computed in canonical units and are therefore comparable across LBC platforms and scanners.

[0231] Slide-level aggregation and classification (Steps 115 / 115A / 115B).

[0232] Per-cell attributes are aggregated over all cells and per class to form the slide-level vector (distribution summaries, top-K order statistics, spatial descriptors: KDE percentiles / peak-to- median, DBSCAN cluster count / size / fraction-in-clusters, nearest-neighbor distances, Moran’s I). The classifier (e.g., CatBoost) operates on these features, and the raw scores are calibrated by isotonic regression to yield a slide-level P(Dysplasia). An operating point may be chosen to maximize automatic exclusion of negative slides subject to a user-specified false-positive-rate bound; automatic exclusions are logged for audit. The QC state (OK / WARNING / ERROR) gates presentation and triage.

[0233] Training, harmonization and validation (Step 118). Training data are iteratively expanded by expert accept / refine / reject of detector proposals across multiple laboratories and scanners. All sites are harmonized by resampling to the canonical MPP and applying stain normalization; augmentations preserve morphology while varying stain / blur within validated limits.

[0234] Exemplary validation: development and site-stratified evaluation on 12 097 LBC slides from 10 laboratories (10 153 ThinPrep, 1 944 SurePath) produced whole-slide classification AUC - 0.93; platform-aware QC bands were tuned accordingly.

[0235] Graphical User Interface (GUI) for Final Diagnosis Presentation

[0236] Fig. 4 illustrates the Graphical User Interface (GUI) designed to present the final diagnosis of a PAP smear analysis using whole-slide imaging (WSI). The GUI 400 is structured to provide a clear, detailed, and user-friendly visualization of the diagnostic results, facilitating efficient and accurate interpretation by pathologists and clinicians.

[0237] The interface comprises (i) a pan-and-zoom WSI viewing pane 410 that renders the slide at the canonical spatial resolution with optional overlays for detected instances and clusters, and (ii) a gallery matrix 420 of enlarged fragments 16 (per-cell thumbnails) selected from regions identified as containing dysplastic or otherwise relevant cells. Selecting a thumbnail 16 recenters the WSI viewing pane 410 on the corresponding coordinates; conversely, navigating the WSI pane can filter the gallery to thumbnails whose coordinates lie within the current viewport.

[0238] Each fragment 16 within the matrix 420 is labeled with a detected cell type 421 (for example, LSIL, HSIL, AGC-NOS, AIS), and is accompanied by a confidence indicator 422 indicating the confidence that the cell is dysplastic. Visual emphasis is provided by an outline 423 or background tint that may be dependent on the cell type or confidence band. When the user hovers the cursor over a fragment 16, a tooltip 424 appears showing additional metadata, such as the confidence value, instance dimensions in canonical units, and other per-cell features.

[0239] At the top, the GUI provides a set of controls 430 for efficient triage and review. These include filters 431 that allow selection by cell class, limiting the gallery to items within the current WSI viewport, and displaying per-class top-K thumbnails. A scrollbar 432 adjusts a confidence threshold so that only thumbnails with confidence above the selected value are shown. Sorting options allow ordering by class, confidence, or spatial order. A set of brightness / contrast / gamma adjustment buttons 433 allow adapting the display specifics of the image. The controls also include buttons to control display of annotations, such as to display all annotations (434), visible only (435) or top 5 (436). A set of scaling buttons 437 allows scaling of image, boxes or margin. A dedicated button 439 indicates submission of diagnosis. The gallery supports per-thumbnail accept / reject actions 425. Each accept (v) or reject (x) emits a structured label event that triggers an immediate, lightweight recalculation of the slide-level risk score, thereby updating the score and the worklist triage order in real-time without reloading the WSI. Accepted / rejected thumbnails are visually flagged, and their corresponding regions are marked in the WSI pane to maintain context.

[0240] The GUI surface also presents the calibrated Probability of dysplasia (P(Dysplasia)) for the whole slide, together with a compact QC banner that reflects the current QC state (OK / WARNING / ERROR). When the state is WARNING, confidence displays may be attenuated and the case priority lowered; when ERROR, automated results are blocked and reacquisition guidance is shown, while navigation and manual review remain available.

[0241] By integrating a pan-and-zoom WSI view with the gallery matrix 420, confidence indicators 422, class labels 421, outlines 423, tooltips 424, and operator controls 430-439, the GUI provides an interactive environment in which users can rapidly review, filter and confirm findings, and in which user confirmations feed back into the diagnostic pipeline to refine the slide-level assessment and worklist prioritization.

[0242] Fig. 5 illustrates a functional diagram of the diagnostic analysis system for cytology microscopic images. The diagram shows the cooperating software modules and data flows that implement the method steps, including scale harmonization, stain normalization, multi-stage quality control, object detection, per-cell feature extraction, slide-level aggregation, calibrated classification, user interaction and audit. The system may be realized as a standalone computer or as a network application operated by cloud computing resources.

[0243] The system comprises a diagnostic module 510 connected to external sources such as a scanner 501 and an image database 502, and to user VO devices including keyboard 521, mouse 522 and display 523. Within the diagnostic module 510 the following components are provided.

[0244] An input image interface 511 is configured to execute Step 110. The interface receives whole-slide images (WSIs) 10, reads scanner micrometers-per-pixel (MPP) metadata and stores a transform Traw^canonical. It harmonizes each WSI to the canonical spatial resolution (e.g., -0.450 pm / px at 20*) by resampling (bicubic) and normalizes photometric interpretation. The harmonized WSI and metadata are forwarded downstream.

[0245] A preprocessing module 511 A is configured to execute Step 110A. This module performs stain normalization (e.g., Macenko, with Reinhard / Vahadane alternatives) and foreground refinement in HSV / OD space, including debris / mucus suppression by small-object removal and morphological smoothing. The resulting stain-normalized image and foreground mask are supplied to QC and tiling.

[0246] A whole-slide quality control module 512 is configured to execute Step 111. The module computes slide-level metrics (brightness, colorfulness, focus / contrast including RMS contrast, variance of Laplacian and Tenengrad, stain-range conformance, foreground / tissue coverage and cellularity proxies) at the canonical resolution, compares them to quantile bands and assigns an OK / WARNING / ERROR state that governs downstream triage and acceptability. Platform-aware bands (e.g., ThinPrep tighter focus / density; SurePath slightly wider focus) may be selected from configuration.

[0247] A tile segmentation module 513 is configured to execute Step 112. It divides the harmonized WSI 10 into overlapping tiles 11 with deterministic slide-aligned coordinates, prepares context versions (e.g., 1 / 4 and 1 / 16 views) and skips or deprioritizes tiles with negligible foreground.

[0248] A tile quality control module 514 is configured to execute Step 113. Per-tile metrics (brightness, colorfulness, focus / contrast, stain-range conformance, foreground coverage, colorful-tile test) are measured and compared to quantile bands; tiles are admitted (OK), admitted with caution (WARNING) or rejected (ERROR). QC metadata are attached to each tile.

[0249] An Al detection module 515 is configured to execute Step 114. In one embodiment this is a YOLO-family CNN with multi-scale heads that consumes each accepted tile and its context versions to output per-cell payloads (x, y, w, h, class, confidence) in canonical coordinates. Perclass NMS and confidence-weighted box voting fuse multi-scale proposals.

[0250] A segmentation and feature-extraction module 515A is configured to execute Step 114A. Around each detection the module derives nuclear and cytoplasmic masks (seeded watershed or UNet) and computes morphological and texture features including nuclear area, N / C ratio, circularity, eccentricity, GLCM chromatin-texture measures on a hematoxylin-dominant channel, and local cell density within a 50 pm neighborhood, all in canonical units.

[0251] A prediction aggregation module 516 configured to execute Step 115. It maps detections to slide-level coordinates, resolves duplicates across overlapping tiles by slide-level NMS (per class) with tie-breaks by confidence and tile quality, and produces a consolidated instance set with thumbnails and provenance.

[0252] A slide-level feature aggregation module 516A is configured to execute Step 115A. It converts per-cell attributes into a fixed-length slide-level feature vector by computing distributional summaries (count, mean, SD, median, IQR, p5 / p95, min, max, coarse histograms, top-K order statistics) over all cells and per class, and spatial-distribution descriptors (KDE percentiles and peak-to-median ratios, DBSCAN cluster count / size / fracti on-in-clusters, nearest- neighbor distances, Moran’s I) with appropriate normalization by foreground area or total cell count.

[0253] A classifier and calibration module 516B is configured to execute Step 115B. A gradient- boosted decision-tree classifier (e.g., CatBoost) ingests the slide-level vector and outputs a raw score that is converted to a calibrated Probability of dysplasia P(Dysplasia) by isotonic regression learned on a held-out validation set. An operating threshold may be applied to maximize automatic exclusion of negative slides subject to a configured false-positive-rate bound.

[0254] A diagnosis quality control module 517 is configured to execute Step 116. It performs diagnosis-level checks (expected ranges for normal / abnormal counts, spatial consistency, feature coherence, influence of marginal-quality tiles) and assigns an OK / WARNING / ERROR diagnosis state; ERROR blocks automated output and requests reacquisition or manual review.

[0255] A final diagnosis module 518 is configured to execute Step 117. It synthesizes the calibrated P(Dysplasia), per-class counts, top-K summaries and QC state into a final diagnostic report and triage decision, and logs automatic exclusions for audit.

[0256] A GUI module 519 provides interactive presentation as in Fig. 4. A WSI viewing pane and a gallery of per-cell thumbnails (with class labels and confidence scores) support real-time filters (class, confidence, viewport) and per-thumbnail accept / reject actions that immediately update the slide-level risk score and reorder the worklist without reloading the WSI. User actions are persisted for model improvement.

[0257] For configuration, the versioning and audit the diagnostic module 510 may further comprise or be coupled with: a model store 524 holding detector and classifier weights with version identifiers; a QC configuration store 525 holding quantile bands and platform / site profiles; an audit log 526 recording QC outcomes, operating-point decisions and automatic exclusions; a feedback collector 527 capturing GUI accept / reject events for iterative training.

[0258] Data produced by modules 511-516B are consumed downstream and may be fed back to a training and harmonization service 520 that performs multi-lab data standardization (canonical MPP, stain normalization), site-stratified validation and calibration refresh as described in Step 118.

[0259] Fig. 6 illustrates an embodiment of a hardware configuration 600 suitable for implementing the diagnostic analysis system of Fig. 5. The configuration may be realized as a standalone workstation, an on-premise server, or a cloud-hosted instance; functional blocks can be colocated or distributed without departing from this embodiment.

[0260] The system comprises the following modules.

[0261] A data interface 601 provides connectivity to external data sources, such as a slide scanner 501 and an image database 502. The interface may include high-throughput ports (e.g., USB-3.x) and / or Ethernet for ingesting whole-slide images (WSIs) and associated metadata (including micrometers-per-pixel, MPP).

[0262] A memory 602 (non-transitory computer-readable medium) stores an operating system and application modules whose instructions, when executed by the processor 603 (and, where present, the accelerator 606), cause the system to perform Steps 110-118 as described herein. Memory 602 can also hold runtime data structures including the transform Traw— >canonical, tiling metadata, per-cell payloads and feature vectors, and configuration such as QC quantile bands and calibration mappings.

[0263] A processor 603, such as one or more general-purpose CPUs, is configured to execute the program instructions implementing the pipeline modules 511, 511A, 512-519 and 516A-516B, including resampling to the canonical spatial resolution, stain normalization, quality-control computations, slide-level aggregation and GUI services.

[0264] A graphics / Al accelerator 606 (optional) expedites convolutional inference for the detector (module 515), segmentation and feature extraction (module 515A), and, where applicable, gradient-boosted inference for the classifier (module 516B).

[0265] A mass storage 607 (e.g., SSD / NVMe) for WSIs, derived pyramids at the canonical resolution, thumbnails, audit logs and versioned model artefacts (corresponding to the model store 524, QC configuration store 525 and audit log 526 of Fig. 5).

[0266] A network interface 608 provides LAN / WAN connectivity (Ethemet / Wi-Fi) to remote scanners, PACS / LIS systems, cloud resources and update services. In a cloud embodiment, 608 supports secure transport for model updates and for collecting user feedback events (the feedback collector 527).

[0267] A control interface 604 ports for user input devices such as keyboard 521 and mouse 522, enabling interaction with the GUI module 519.

[0268] A display interface 605 outputs image data to a display 523 for the WSI viewing pane and the gallery of per-cell thumbnails with confidence scores and class labels.

[0269] In operation, WSIs and metadata are received via the data interface 601 or network interface 608 and are written to storage 607. The processor 603 (optionally with accelerator 606) executes code stored in memory 602 to harmonize scale to the canonical micrometers-per-pixel, apply stain normalization and debris suppression, perform whole-slide and tile-level QC, tile the WSI, run object detection, per-cell segmentation and feature extraction, aggregate slide-level features, classify and calibrate the slide-level probability of dysplasia P(Dysplasia), and drive the user interface. Configuration items (e.g., QC bands, platform / site profiles) and versioned model weights are persisted on storage 607 and loaded into memory 602 at runtime. All QC outcomes, operating-point decisions and user accept / reject actions are logged to storage 607 for audit and for iterative model improvement as described in Step 118.

[0270] This hardware configuration supports both single-machine and distributed deployments; modules may be split across multiple nodes connected by the network interface 608 without departing from the scope of this embodiment.

[0271] Fig. 7 schematically illustrates the flow for receiving a whole-slide image (WSI) and harmonizing its spatial scale to a canonical micrometers-per-pixel (MPP) resolution prior to tiling and analysis (Step 110), in a form of sub-modules of the input image interface 511.

[0272] A WSI 10 together with its associated metadata 10M (including scanner-reported MPP, pyramid levels, photometric interpretation and orientation tags) is provided to the input image interface 511. At the input image interface, an MPP resolver 701 reads the native MPP from 10M; if the value is missing or inconsistent, the resolver estimates MPP from one or more of: a scanner profile 702 (holding data for various models of scanners), a scale bar 703 embedded within the WSI, or an anatomical -prior estimator 704 (configured to estimate dimensions based on known biological priors) that compares observed cell sizes to expected physical ranges in cytology. The selected MPP is validated against internal consistency checks; if the residual uncertainty exceeds a predefined bound, a QC notice 705 is emitted for later display in the QC modules.

[0273] Given the native and target scales, a scale calculator 706 computes the affine factor s=MPPnative / MPPcanonical .

[0274] A level selector 707 chooses an appropriate pyramid level and a resampler 708 applies bicubic interpolation (with anti-alias prefiltering for down-sampling) to obtain a harmonized image at the canonical resolution (in one embodiment, 0.450 pm / px (±0.04) at 20x). During resampling, slide orientation and coordinate conventions are normalized to a rightwarddownward system with origin at the upper-left corner.

[0275] A transform registry 709 stores the transformation Traw^canonical (and its inverse) so that tile coordinates, per-cell thumbnails and UI interactions can be mapped losslessly between raw and canonical space. Optionally, a color-space normalizer 710 converts input to a standard color space (e.g., sRGB) and reconciles photometric interpretation so intensity and color channels are consistent across scanners.

[0276] The outputs of this flow are a harmonized WSI IO C at the canonical MPP together with the stored transform Traw^canonical and updated metadata. For reference, Fig. 7 indicates that, at 0.450 pm / px, a 1280 x 1280 px tile used later corresponds to approximately 576 pm x 576 pm. The harmonized image IO C proceeds to stain normalization and debris suppression (Step 110A) and thereafter to whole-slide QC and tiling, ensuring that all subsequent measurements, windows and thresholds are defined in canonical physical units and are comparable across scanners and laboratories.

[0277] Fig. 8 schematically illustrates the per-cell segmentation and feature-extraction pipeline executed after object detection (Step 114), in a form of sub-modules of the segmentation and feature-extraction module 515A. The inputs are the stain-normalized, canonical -resolution tile 11 (Step 110 A) together with the detector’s per-cell payloads 13 (x,y,w,h, class, confidence) mapped to slide coordinates. The pipeline derives nucleus and cytoplasm masks and computes morphological and chromatin-texture features in canonical physical units, together with a neighborhood-based local density measure. The resulting feature records are linked to slide coordinates and QC lineage for downstream aggregation (Steps 115, 115A).

[0278] In one embodiment the flow comprises the following functional elements: detection crop window 801 : around each detection box, a processing window is formed with a small margin (e.g., 10-20% of box size in canonical pixels). stain channel extraction 802: a hematoxylin-dominant channel is computed (OD-space deconvolution) to stabilize chromatin texture across laboratories and scanners. segmentation module 803: within the window, nucleus and cytoplasm are segmented by seeded watershed initialized from the box centerline; in another embodiment a compact UNet is used. masks 804: nuclear and cytoplasmic masks are produced and refined by morphological operations (hole filling, small-object removal, contour smoothing) subject to plausibility bounds in pm2and shape. neighborhood window 807: A circular neighborhood of 50 pm radius (expressed in canonical pixels) is defined around the instance center for local statistics. morphological features 805 : from the masks: nuclear area (pm2), cytoplasmic area (pm2), N / C ratio (AN / AC), circularity (4KAN / P2), and eccentricity (from second central moments). chromatin-texture features 806: on the hematoxylin-dominant channel: GLCM contrast, entropy, homogeneity, computed over fixed windows / offsets in canonical pixels. local cell density 808: count of neighboring detected cells whose centers fall within 50 pm (neighborhood window). plausibility / QC checks 809: feature values are checked against validated physical ranges and mask-quality indicators; out-of-range values are flagged and, where feasible, segmentation is re-run with conservative parameters. coordinate & provenance assembler 811 : slide-level coordinates are attached (tile origin + local offset), together with detector class / confidence, source tile ID and its QC state, and references to per-cell thumbnails (fragment 16). feature record output 812: a structured per-cell feature record is emitted for aggregation, comprising the attributes above and flags for any imputations or cautions.

[0279] Because all measurements are performed at the canonical spatial resolution and on stain- normalized content, the features are comparable across scanners, laboratories and liquid-based cytology (LBC) platforms. The output records feed consolidation (Step 115) and slide-level feature aggregation (Step 115 A), and support interactive review in the GUI (Fig. 4).

[0280] Fig. 9 - Slide-level feature aggregation and spatial-distribution descriptors

[0281] Fig. 9 schematically illustrates the formation of a fixed-length slide-level feature vector from the consolidated detections (Step 115) and the per-cell feature records (Step 114A), in a form of sub-modules of the slide-level feature aggregation module 516A. All quantities are expressed at the canonical spatial resolution, and, where appropriate, are normalized by foreground tissue area or total cell count to ensure comparability across scanners, laboratories and preparations.

[0282] Inputs: the featurizer ingests (i) the consolidated per-cell payloads (x,y,w,h, class, confidence) in slide-level canonical coordinates with tile / QC lineage, (ii) the per- cell morphological and chromatin-texture features (areas, N / C, circularity, eccentricity, GLCM measures, local density within 50 pm), and (iii) foreground / tissue masks and slide metadata.

[0283] Processing overview: a gating stage 901 applies a low-confidence floor and optional class smoothing; cells are grouped over all detections and per cytology class (e.g., LSIL, HSIL, SCC, AGC-NOS, AGC-NEO, AIS, ADCA, Endocervical). For each numeric per-cell attribute, a panel of distributional summaries is computed 902: count, mean, standard deviation, median, IQR, p5 / p95, min, max, together with coarse histograms and top-K order statistics (e.g., mean / min over the highest-scoring K cells, KG{5, 10,50}). In parallel, spatial-distribution descriptors are derived 903 from coordinates of all cells and of abnormal subsets: a kernel-density estimate (KDE) with density percentiles and peak-to-median ratios (bandwidth in pm), DBSCAN clustering (s in pm; minPts set empirically) yielding cluster count, average / max cluster size and fraction in clusters, nearest-neighbor distance statistics (mean, median, p5 / p95) and a Moran’s I spatial autocorrelation score.

[0284] Features are then normalized 904 (e.g., by foreground area or total cell count). Missing values (e.g., for absent classes) receive neutral imputation and presence flags. Where enabled, contributions from tiles marked WARNING by Step 113 may be down-weighted. Finally, all summaries are concatenated 905 in a fixed order to form the slide-level feature vector, accompanied by a compact descriptor map that records feature indices. The vector is passed to the classifier and calibration stage (Step 115B).

[0285] Outputs: (1) The slide-level feature vector and associated metadata (normalization, imputation flags, down-weighting notes); (2) Lightweight summaries (counts per class, KDE hotspot indicators, cluster counts) for UI preview and audit.

[0286] Fig. 10 schematically illustrates the slide-level classification and calibration pipeline (Step 115B), in a form of sub-modules of the classifier and calibration module 516B. The pipeline ingests the slide-level feature vector from Step 115A, produces a calibrated Probability of dysplasia P(Dysplasia) and applies an operating threshold to support triage and optional automatic exclusion of likely negative slides, with behavior gated by the QC state from Steps 111-116.

[0287] Inputs: the classifier receives (i) the slide-level feature vector with normalization / imputation flags, (ii) the slide’s QC state (OK / WARNING / ERROR), and (iii) configuration records containing the model weights and calibration mapping.

[0288] Processing overview: a lightweight pre-processor 1001 confirms feature ordering and applies the same robust scaling as in training. The vector is then fed to a gradient-boosted decision-tree classifier 1002 (e.g., CatBoost) that outputs a raw score. An isotonic-regression calibration mapping, learned on a held-out validation set during training, converts 1003 the raw score to a calibrated probability P(Dysplasia). An operating-point selector 1004 compares the probability to a threshold chosen to maximize automatic exclusion of negative slides, subject to a user-specified false-positive-rate (FPR) bound. The selector emits a risk band and a triage decision (e.g., auto-exclude vs. review), and logs the decision for audit. The QC gate enforces policy: ERROR blocks automated diagnosis; WARNING attenuates per-cell confidence displays and lowers priority; OK proceeds normally.

[0289] Outputs: (1) Calibrated P(Dysplasia) and risk band for the GUI (Fig. 4); (2) Triage decision and updated worklist order; (3) Audit records of the operating-point decision and the versions of the model and calibration mapping used.

[0290] During model development, site-stratified cross-validation is used to train the boosted trees; a held-out set fits the isotonic calibration. The resulting model weights and calibration mapping are versioned and deployed with the runtime.

[0291] Finally, the following section summarizes the QC policy that governs processing and presentation at three levels of the pipeline: whole-slide QC (Step 111), tile-level QC (Step 113) and diagnosis-level QC (Step 116). Each level measures a fixed panel of metrics in canonical units (e.g., brightness, colorfulness, focus / contrast including RMS contrast, variance of Laplacian and Tenengrad, stain-range conformance, foreground / tissue coverage, cellularity proxies and, at diagnosis level, normal / abnormal counts and spatial-consistency checks). Metrics are compared to quantile bands learned from a multi-lab training corpus to assign one of three QC states:

[0292] OK: all metrics within central bands; automated diagnosis proceeds normally.

[0293] WARNING: one or more metrics outside central bands but within acceptance; a caution indicator is surfaced, confidence displays are attenuated and case priority is reduced; automated diagnosis continues.

[0294] ERROR: one or more metrics outside acceptance; automated diagnosis is blocked and reacquisition or manual review is requested.

[0295] Whole-slide QC (Step 111). The WSLlevel metrics are computed after harmonization (Step 110) and stain normalization (Step 110A). Bands are site / preparation aware: for LBC, ThinPrep uses tighter density / focus bands, while SurePath permits a slightly wider focus band. The WSI QC state (OK / WARNINGZERROR) is propagated downstream and influences tiling priority, later diagnosis QC and GUI presentation.

[0296] Tile-level QC (Step 113). Each tile 11 inherits the foreground mask and undergoes pertile QC using the same banding policy. Tiles are admitted (OK), admitted with caution (WARNING, later down-weighted) or rejected (ERROR). If a configurable fraction of tiles are rejected or only WARNING tiles remain, a slide-level escalation is raised (WARNINGZERROR) and recorded.

[0297] Diagnosis-level QC (Step 116). After consolidation and feature aggregation (Steps 115, 115A), the system checks abnormal / normal counts against expected ranges, spatial descriptors (KDE, DBSCAN, nearest-neighbor, Moran’s I) and feature coherence with the detected classes. The diagnosis QC state gates the classifier / calibration output (Step 115B): ERROR blocks automated diagnosis regardless of the raw probability; WARNING attenuates per-cell confidence display and lowers priority; OK proceeds normally. All QC decisions, thresholds and actions are logged to an audit trail together with the versions of the bands configuration and model / calibration used. A monitored feedback loop may tune WARNING bands within validated ranges over time; acceptance bands remain fixed unless revalidated. While the invention has been described with respect to a limited number of embodiments, it will be appreciated that many variations, modifications and other applications of the invention may be made. Therefore, the claimed invention as recited in the claims that follow is not limited to the embodiments described herein.

Claims

1. CLAIMS1. A computer-implemented method for assisting diagnosis of dysplasia from a liquid-based cytology whole-slide image (WSI) (10), by means of a multi-stage pipeline that combines deeplearning-based detection with engineered feature extraction and a calibrated machine-learning classifier, the method comprising:(a) receiving (110), by an input image interface (511), a digital WSI of a cytology specimen and resampling the WSI to a canonical spatial resolution (b) detecting (114), by an object-detection neural network (300), individual cell (12) instances and outputting, for each detected cell, a predicted class, a confidence value and image coordinates as a per-cell payload (13);(c) computing (114A), for each detected cell and a predefined neighborhood thereof at the canonical resolution, a plurality of morphological and texture features;(d) aggregating (115 A) the per-cell features across the WSI to obtain a slide-level feature vector comprising statistical measures and spatial-distribution descriptors;(e) inputting (115B) the slide-level feature vector to a gradient-boosted decision-tree classifier and calibrating an output thereof to produce a calibrated slide-level probability of dysplasia;(f) computing (111) slide-quality metrics from the WSI at the canonical resolution and, by comparing each metric to pre-configured quantile bands derived from a training distribution, assigning a quality state and controlling downstream operation including at least triage priority and acceptability of Al diagnosis;(g) presenting (117), via a graphical user interface (400), the WSI and Al results including a gallery (420) of per-cell thumbnails (16) annotated with respective detector confidence scores (422), derived from the confidence value output by the object-detection neural network, the graphical user interface being configured to accept per-thumbnail user confirmations or rejections and, in response, to update the slide-level probability of dysplasia and a worklist triage order without reloading the WSI (10); and(h) iteratively expanding and harmonizing training data (118) by proposing cell annotations with the object-detection neural network on additional slides, incorporating expert accept / reject decisions, and standardizing images from multiple laboratories by resampling to a common micrometers-per-pixel resolution and applying stain normalization, for training the objectdetection neural network and the classifier.

2. The method of claim 1, wherein the morphological and texture features comprise at least nuclear area, nucleus-to-cytoplasm ratio, one or more shape descriptors including circularity andeccentricity, one or more chromatin texture descriptors, and a local cell-density measure, and wherein the slide-level feature vector comprises, for each feature, one or more of: mean, median, variance, inter-quartile range, minimum, maximum, specified percentiles, and spatial-distribution descriptors.

3. The method of any of previous claims, wherein the slide-quality metrics include at least a celldensity metric and a focus or contrast metric, and wherein assigning the quality state comprises issuing a WARNING when a metric lies outside a first quantile band and an ERROR when outside a diagnostic-acceptability band, a WARNING causing surfacing of a caution indicator, modification of confidence display and lowering of slide-processing priority, and an ERROR blocking Al diagnosis and requesting reacquisition; the method further comprising selecting a classifier threshold to maximize a proportion of negative slides automatically excluded from review subject to a user-specified false-positive-rate bound, and re-ordering the worklist accordingly while logging automatically excluded cases for audit.

4. The method of any of previous claims, wherein iteratively expanding and harmonizing the training data comprises: manually annotating at least a predetermined number (N) of cells per dysplasia class on initial slides; training the object-detection neural network; using the trained object-detection neural network to propose cell annotations on additional slides; expert accepting, refining, or rejecting proposals and adding accepted annotations and rejected false positives to the training set; repeating to accumulate a verified multi-lab dataset; resampling slides to the common resolution; applying stain-normalization; and training the object-detection neural network and the classifier using focal loss and minority-class oversampling.

5. The method of any of previous claims, wherein the graphical user interface (400) renders, in a first pane, a pan-and-zoom view (410) of the WSI (10) and, in a second pane, the gallery (420) of per-cell thumbnails (16) with confidence indicators (422), selection of a thumbnail (16) recentering the first pane (410) at corresponding coordinates, the gallery (420) being filterable in real time to thumbnails whose coordinates lie within a current viewport of the first pane (410).

6. The method of claim 5, further comprising a class-quota input to display per-class top-K thumbnails and a confidence-threshold input to hide thumbnails below a selected confidence, each accept or reject action (425) emitting a structured label event that immediately updates the slide-level probability of dysplasia and the worklist ranking and is persisted for model retraining.

7. The method of any of claims 5-6, wherein the graphical user interface further exposes (i) a class-quota input to display per-class top-K thumbnails and (ii) a confidence-threshold input, and wherein each per-thumbnail accept or reject action emits a structured label event that immediately updates the slide-level probability of dysplasia and reorders a worklist without reloading the WSI.

8. The method of any of previous claims, wherein stain normalization is performed by a Macenko-type optical-density procedure, and a foreground / tissue mask is derived in HSV and / or OD space with small-object removal to suppress debris and mucus.

9. The method of any of previous claims, wherein the WSI is divided into fixed-size tiles with an overlap, and a scheduler is configured to prioritize tiles with higher foreground coverage.

10. The method of any of previous claims, wherein, in addition to a highest-magnification tile input, one or more context versions are generated that preserve the detector input size while enlarging field-of-view, including at least a 1 / 4-scale or a 1 / 16-scale context in which the original tile occupies the central portion.

11. The method of any of previous claims, wherein duplicate detections are suppressed by per- class non-maximum suppression at an intersection-over-union and multi-scale proposals are fused by confidence-weighted box voting, and slide-level de-duplication across overlapping tiles prefers instances originating from tiles with higher quality-control scores.

12. The method of any of previous claims, wherein per-cell segmentation is performed within an enlarged detection window and comprises seeded watershed on a hematoxylin-dominant channel or a compact UNet.

13. The method of any of previous claims, wherein aggregating comprises computing, for each numeric per-cell attribute, top-K order statistics and spatial-distribution descriptors including kernel-density percentiles and peak-to-median ratios, DBSCAN cluster count / size / fracti on-in- clusters with s defined in pm and site-validated minPts, nearest-neighbor distance statistics, and a spatial autocorrelation score, the features being normalized by foreground area and / or total detected cell count and contributions from WARNING-state tiles being down-weighted.

14. A non-transitory computer-readable storage medium storing instructions which, when executed by one or more processors, cause performance of the method of any one of previous claims.

15. A system for assisting diagnosis of dysplasia from a liquid-based cytology whole-slide image (WSI) (10), comprising at least one processor and a non-transitory memory storing instructions which, when executed by the at least one processor, cause the system to implement modules configured to: an input image interface (511) configured to receive (110) a digital WSI of a cytology specimen and resample the WSI to a canonical spatial resolution; an object-detection neural network (300) configured to detect (114) individual cell (12) instances and, for each detected cell, output a per-cell payload (13) comprising a predicted class, a confidence value and image coordinates; a feature-extraction module (515A) configured to compute (114A), for each detected cell and a predefined neighborhood at the canonical resolution, a plurality of morphological and texture features; an aggregation module (516 A) configured to aggregate (115 A) the per-cell features across the WSI to obtain a slide-level feature vector comprising statistical measures and spatial - distribution descriptors; a classifier and calibration module (516B) configured to input (115B) the slide-level feature vector to a gradient-boosted decision-tree classifier and calibrate an output thereof to produce a slide-level probability of dysplasia; a quality-control module (512) configured to compute (111) slide-quality metrics from the WSI at the canonical resolution and, by comparing each metric to percentile thresholds derived from a training distribution, assign a quality state and control downstream operation including at least triage priority and acceptability of Al diagnosis; a graphical user interface (400) configured to present (117) the WSI and Al results including a gallery (420) of per-cell thumbnails (16) annotated with respective confidence scores derived from the confidence value output by the object-detection neural network, the interface being further configured to accept per-thumbnail user confirmations or rejections and, in response, to update the slide-level probability of dysplasia and a worklist triage order without reloading the WSI (10); and a training-data expansion and harmonization module (520) configured to iteratively expand and harmonize training data (118) by proposing cell annotations with the object-detectionneural network on additional slides, incorporating expert accept / reject decisions, and standardizing images from multiple laboratories by resampling to a common micrometers- per-pixel resolution and applying stain normalization, for training the object-detection neural network and the classifier.

Images