Artificial intelligence based platform for automated pap smear slide analysis and pathological counselling

Abstract

The invention provides an artificial-intelligence–based platform for automated analysis of Pap smear slides. The system integrates preprocessing, dual deep learning models (CNN-Transformer and EfficientNet-LSTM), and reporting modules to classify cytological images into diagnostic categories including NILM, ASC-US, LSIL, HSIL, and SCC. The platform includes interpretability outputs, secure data management, and active learning capabilities, enabling scalable, accurate, and transparent cervical-cancer screening across diverse laboratory settings. The platform is suitable for implementation in clinical pathology laboratories, telemedicine networks, and population-level screening programs. It allows rapid, reproducible, and objective evaluation of cytology slides, improving throughput while reducing reliance on manual slide examination. The invention is deployable on local servers, private clouds, or public cloud infrastructures and is adaptable to various cytology specimens. Adoption of this system enhances early detection of precancerous lesions, reduces inter-observer variability, and facilitates integration with laboratory information systems. It also supports continuous learning and adaptation to evolving imaging protocols. The invention provides multiple advantages over existing methods: (1) Dual-model architecture combining CNN-Transformer and EfficientNet- LSTM ensures robust classification across both dense and sparse cellular regions. (2)Preprocessing and adaptive tiling improve feature extraction from variable-quality images. (3)Interpretability modules enhance clinician trust and facilitate regulatory compliance. (4) Active learning allows continuous improvement based on expert feedback. (5) Modular deployment supports both local laboratory and cloud-based operation. (6) Data security features meet international standards for privacy and encryption. (7) High scalability enables screening of large slide volumes without compromising accuracy. The invention provides a clinically relevant, efficient, and technologically advanced solution for automated cervical-cancer screening, applicable in diverse healthcare environments and adaptable to future cytology domains.

Description

[0001] Artificial Intelligence Based Platform for Automated Pap Smear Slide Analysis and pathological counselling

[0002] TECHNICAL FIELD

[0003] The present invention relates generally to the application of artificial-intelligence systems in digital cytopathology and, more particularly, to a computer-implemented platform that performs automated analysis of Pap smear slides for detection and grading of cervical epithelial abnormalities. The invention lies at the intersection of computer vision, machine learning, and clinical diagnostics, and provides an end-to-end environment for image acquisition, preprocessing, feature extraction, classification, and automated reporting. It further concerns cloud-based data management and model-driven decision support systems that enhance the accuracy, speed, and scalability of cervical-cancer screening workflows. Specifically, the invention provides a hybrid artificial-intelligence architecture combining convolutional-neural-network (CNN) and Transformer components for spatial-contextual feature learning, and an alternative EfficientNet- LSTM configuration for temporal or sequential cytology-image interpretation. These architectures are integrated within a modular, web-accessible platform enabling both laboratory and telemedicine deployment. The invention aims to minimize human subjectivity in cytological interpretation and to extend diagnostic capabilities to low-resource or remote settings.

[0004] BACKGROUND

[0005] Cervical cancer remains a major public -health concern worldwide. The Papanicolaou test (Pap smear) is one of the most successful screening methods developed to detect precancerous and cancerous changes in the cervix. Traditionally, cytotechnologists and pathologists manually examine glass slides under a microscope to identify morphological abnormalities in squamous or glandular epithelial cells. While effective, manual screening is inherently limited by inter-observer variability, fatigue, and time consumption. Large-scale population screening further strains the available workforce and laboratory infrastructure.

[0006] Recent advancements in digital microscopy and whole-slide imaging (WSI) have enabled conversion of glass slides into high-resolution digital images. These digital slides open the opportunity for computational image analysis and artificial-intelligence-based classification. Several research prototypes have been proposed to automatically detect abnormal cervical cells using handcrafted features such as shape, color, and texture descriptors combined with conventional classifiers like support-vector machines (SVMs) or random forests. However, these methods require extensive feature engineering, are sensitive to variations in staining and focus, and often fail to generalize across datasets collected from different laboratories or imaging devices. Deep-learning methodologies, particularly CNNs, have shown superior performance in medicalimage analysis, including histopathology and radiology. Nevertheless, the direct application of CNNs to Pap smear images faces unique challenges. Cytology slides contain sparse cellular components scattered over large backgrounds, variable staining intensity, overlapping cells, and artifacts such as mucus or blood. Moreover, the diagnostic label for a slide depends on the distribution and proportion of abnormal cells, not merely on the presence of a single atypical nucleus. Consequently, a robust diagnostic system must capture both local cellular morphology and global spatial context. Existing commercial solutions for automated Pap-smear screening, including certain proprietary systems developed by large diagnostic companies, often rely on hardware-specific scanners and closed algorithms that require expensive licensing. Such systems may achieve high sensitivity but are cost-prohibitive and lack interoperability with heterogeneous laboratory information systems. Furthermore, they typically operate as black boxes, providing limited transparency regarding the basis of classification decisions, which hinders regulatory approval and clinician trust.

[0007] From a clinical perspective, the classification schema recommended by the Bethesda System (TBS) categorizes Pap smear results into several diagnostic groups: Negative for Intraepithelial Eesion or Malignancy (NIEM), Atypical Squamous Cells of Undetermined Significance (ASCUS), Eow-Grade Squamous Intraepithelial Eesion (LSIL), High-Grade Squamous Intraepithelial Lesion (HSIL), and Squamous Cell Carcinoma (SCC). A reliable automated system must accurately reproduce these categories while maintaining explainability and traceability of its predictions. Despite many academic studies, there is no widely adopted open, standardized, AL based pipeline that performs end-to-end classification across this spectrum.

[0008] In addition to diagnostic accuracy, another challenge lies in the diversity of image sources. Laboratories use different microscopes, cameras, staining protocols, and magnifications. The resulting images vary significantly in resolution, color profile, and focus depth. A robust algorithm must therefore incorporate normalization, stain correction, and data-augmentation techniques to achieve cross-laboratory generalization. Moreover, datasets containing whole-slide images (WSI) can reach gigapixel sizes, requiring efficient tiling, patch extraction, and distributed computation. Cloud computing and web-based medical platforms have emerged to address scalability and accessibility issues in medical AL However, many existing frameworks are limited to specific disease domains, such as dermatology or radiology, and lack optimized workflows for cytopathology. Pap smear analysis presents distinct computational requirements, including cellsegmentation algorithms, morphological-feature quantification, and interpretability modules designed for cell clusters rather than tissue sections. To date, few end-to-end platforms have integrated these elements into a single unified pipeline suitable for routine clinical use.

[0009] The absence of such an integrated platform contributes to delays in diagnosis, inconsistent reporting, and missed opportunities for early intervention. Manual review of every slide by human experts restricts the throughput of screening programs, particularly in regions with limited cytotechnologists. Consequently, many laboratories perform only partial sampling, which increases false-negative rates. An intelligent, automated, and scalable system could reduce human workload, prioritize suspicious slides for expert review, and enhance overall screening coverage. Beyond efficiency, automated systems can also improve objectivity. Unlike human observers, algorithms do not suffer from fatigue or bias and can maintain consistent performance over time. By quantifying morphological parameters such as nuclear-to-cytoplasmic ratio, chromatin texture, and irregular nuclear membrane, machine-learning models can provide standardized criteria for lesion grading. Furthermore, by maintaining comprehensive digital records, Al platforms enable longitudinal tracking of patient samples and facilitate population-level analytics.

[0010] Despite these advantages, several obstacles hinder widespread adoption. Many Al studies in cytology use small, non-representative datasets that lack pathological diversity. Reproducibility across institutions is rarely demonstrated. Moreover, regulatory agencies such as the FDA or EMA require explainable outputs, yet most deep-learning models remain opaque. Existing open-source projects often focus on single-cell segmentation rather than complete diagnostic classification, leaving a gap for a clinically validated, regulatory-compliant, Al-based Pap-smear analysis platform.

[0011] Accordingly, there exists a pressing need for an invention that provides: an end-to-end digitalcytology platform that ingests images from both WSI scanners and conventional microscope cameras; robust Al models capable of detecting and classifying cervical epithelial abnormalities across multiple grading systems; a modular and scalable architecture deploy able on local servers or cloud environments; and integrated reporting and visualization modules aligned with clinical standards such as the Bethesda System.

[0012] The present invention addresses these unmet needs through a hybrid Al-powered system combining advanced image -processing pipelines with deep-learning architectures optimized for cytological morphology. The invention leverages cloud connectivity, secure data storage, and explainable- Al modules to provide reliable and transparent diagnostic support to cytopathologists worldwide.

[0013] SUMMARY

[0014] The present invention provides a comprehensive, Al-driven software platform for automated analysis of Pap-smear slides. The platform integrates data acquisition, image preprocessing, feature extraction, classification, and diagnostic reporting into a single workflow. It is designed to operate with heterogeneous imaging sources, including whole-slide scanners and conventional optical-microscope cameras, without dependence on specific hardware vendors.

[0015] At the core of the system lies a hybrid deep-learning architecture composed of two alternative and complementary models: (a) a CNN + Transformer hybrid optimized for spatial and contextual understanding of cytology images, and (b) an EfficientNet + LSTM network for modeling sequential and temporal relationships between cell clusters captured across different fields of view. Both models are trained using large, annotated datasets of Pap-smear images labeled according to the Bethesda System diagnostic categories.

[0016] The invention further provides a preprocessing subsystem that standardizes input images through color normalization, noise reduction, cell-segmentation refinement, and patch extraction. Each image is divided into overlapping tiles that are dynamically sampled based on cellular density, ensuring that diagnostically relevant regions are prioritized for analysis.

[0017] The classification subsystem outputs probabilistic predictions across five diagnostic classes- NILM, ASC-US, LSIL, HSIL, and SCC, and flags any region with high abnormality confidence for expert review. The system also produces human-readable heatmaps showing regions that most strongly influenced the model’s decision, thereby enhancing interpretability and regulatory compliance.

[0018] Unlike conventional automated cytology devices that rely on proprietary scanners, the present invention is software-centric and deployable in diverse laboratory environments. Its modular architecture supports both offline and online processing modes, enabling standalone use within local pathology labs or centralized cloud-based screening for remote clinics.

[0019] Key Innovations: The invention introduces several distinctive innovations: dual-model Al core combining CNN-Transformer and EfficientNet-LSTM for complementary spatial-temporal analysis; adaptive tiling engine dynamically adjusting tile size and overlap based on cellular content to reduce computational load; explainable Al module providing saliency maps, Grad-CAM visualizations, and textual reasoning summaries; integrated report generator automatically producing standardized diagnostic reports consistent with TBS categories; secure data pipeline incorporating encryption, anonymization, and audit-trail management for regulatory compliance; and continuous learning framework enabling periodic model updates through active-learning loops fed by expert corrections.

[0020] Collectively, these innovations yield an intelligent diagnostic assistant that improves accuracy, consistency, and throughput of Pap-smear screening while maintaining clinical transparency and cross-laboratory adaptability.

[0021] DETAILED DESCRIPTION OF THE INVENTION

[0022] The system comprises five primary modules interconnected via an internal message bus and application-programming interfaces (APIs):

[0023] 1. Image Acquisition and Ingestion Module (IAIM)

[0024] 2. Preprocessing and Normalization Module (PNM)

[0025] 3. Al Inference Core (AIC)

[0026] 4. Reporting and Visualization Module (RVM)

[0027] 5. Data Security and Management Module (DSMM)

[0028] Each module may be implemented as a microservice deployable on local or cloud infrastructure. Communication between modules occurs through encrypted channels using standardized data- exchange formats such as DICOM, OME-TIFF, or ISON-based metadata.

[0029] Image Acquisition and Ingestion Module

[0030] The IAIM interfaces with microscope cameras, slide scanners, or mobile devices. It supports both streaming and batch-upload modes. Upon receiving an image, the IAIM automatically extracts metadata such as resolution, magnification, and staining protocol. The images are then queued for preprocessing. The system incorporates an optional quality-control subroutine that detects out-of- focus regions, improper illumination, or debris contamination, prompting the user to rescan if necessary.

[0031] In low-resource environments, the IAIM can operate with simplified imaging devices, accepting JPEG or PNG files captured through smartphone-assisted microscopy. The module applies geometric correction to mitigate lens distortion and aligns multiple focal planes using focusstacking algorithms to approximate whole-slide quality.

[0032] Preprocessing and Normalization Module

[0033] The PNM performs several sequential operations to prepare raw images for Al analysis:

[0034] 1. Color Normalization: standardizing hematoxylin-eosin or Papanicolaou stain variations through adaptive histogram matching.

[0035] 2. Noise Reduction: applying Gaussian or median filtering to suppress background noise while preserving nuclear boundaries.

[0036] 3. Cell Segmentation: leveraging a lightweight U-Net or Mask R-CNN model trained to delineate individual nuclei and cytoplasm regions.

[0037] 4. Patch Extraction: dividing slides into square tiles (e.g., 256 x 256 pixels) with adaptive overlap to capture adequate context. 5. Data Augmentation: generating synthetic variations via rotation, flipping, and color jitter to improve model robustness.

[0038] After preprocessing, each tile is assigned a unique identifier and stored in a local or cloud database together with its metadata and transformation parameters, ensuring reproducibility.

[0039] Al Inference Core

[0040] The AIC contains two interchangeable model pipelines:

[0041] (1) CNN + Transformer Hybrid

[0042] This pipeline first passes each tile through convolutional layers to extract low-level morphological features such as nuclear texture and cytoplasmic granularity. The resulting feature maps are then flattened and processed by multi-head self-attention layers that capture long-range spatial relationships between cell clusters. The hybrid network produces a vector representation encoding both local and contextual information.

[0043] (2) EfficientNet + LSTM Pipeline

[0044] In this configuration, the EfficientNet backbone serves as a feature extractor for a sequence of image tiles representing contiguous fields of view. The extracted feature vectors are fed into a bidirectional LSTM that models spatial-temporal dependencies, reflecting how cytopathologists scan slides sequentially. The LSTM output is passed to fully connected layers producing final classification probabilities.

[0045] Both pipelines share a common training strategy involving balanced sampling, weighted crossentropy loss to compensate for class imbalance, and early stopping based on validation AUC. The system stores model weights and configuration files within a version-controlled registry, enabling rollback and reproducibility.

[0046] Reporting and Visualization Module

[0047] The RVM aggregates the tile-level predictions into slide-level decisions through majority voting or attention-based pooling. The module then generates a structured report containing:

[0048] The predicted diagnostic category (NILM, ASC-US, LSIL, HSIL, SCC);

[0049] Confidence scores and visual overlays of abnormal regions;

[0050] Quantitative cytomorphometric indices (e.g., nuclear-cytoplasmic ratio, chromatin heterogeneity); Optional narrative explanations derived from the attention weights.

[0051] The report can be exported in PDF or integrated directly into laboratory information systems (LIS).

[0052] Data Security and Management Module

[0053] The DSMM ensures compliance with data-protection regulations such as GDPR, HIPAA, and regional privacy standards. It implements:

[0054] Encryption: AES-256 encryption of images in transit and at rest;

[0055] Anonymization: automatic removal of patient identifiers before processing;

[0056] Access Control: role-based authentication and audit logging;

[0057] Traceability: maintaining a full record of preprocessing, inference, and reporting steps. The module also manages dataset partitioning, model-update scheduling, and synchronization between edge devices and the central server. Secure APIs allow integration with hospital information systems while preserving patient confidentiality.

[0058] Scalability and Deployment

[0059] The architecture is containerized using technologies such as Docker and orchestrated by Kubernetes. This enables horizontal scaling of inference workers according to workload. GPU acceleration is supported for both training and inference stages. The system can be deployed as:

[0060] 1. On-premises installation: suitable for hospitals requiring full data control;

[0061] 2. Private cloud: for networked laboratories within a healthcare group;

[0062] 3. Public cloud service: offering screening-as-a-service to remote clinics.

[0063] The modularity allows future extension to other cytology domains such as urine, sputum, or fine- needle-aspiration smears without architectural redesign.

[0064] Human-in-the-Loop Workflow

[0065] The invention supports a semi-automated review process in which the Al system pre-screens all slides and ranks them by abnormality probability. Cytotechnologists then focus their attention on the highest-risk slides, significantly reducing manual workload while maintaining human oversight. Any corrections provided by experts are fed back into the active-learning subsystem to continually refine model performance.

[0066] CNN-Transformer Hybrid Architecture

[0067] The invention introduces a hybrid deep learning architecture combining the strengths of convolutional neural networks (CNNs) for local texture analysis and transformer encoders for global contextual reasoning.

[0068] The CNN module consists of several convolutional blocks, each including:

[0069] 2D convolution with kernel size 3x3

[0070] Batch normalization

[0071] ReLU activation

[0072] Max pooling

[0073] The transformer module operates on the flattened feature maps from the CNN backbone. The architecture includes multi-head self-attention layers that model spatial relationships among cell clusters and nuclei distributions.

[0074] The hybrid model improves robustness in recognizing subtle inter-class variations between atypical squamous cells and high-grade lesions.

[0075] Pytorch Implementation (Core Code Example)

[0076] Below is a simplified version of the model implemented using Py Torch: import torch import torch, nn as nn import torch, nn. functional as F class CNNTransformerPap(nn.Module): def init (self, num_classes=5, embed_dim=256, num_heads=8, depth=4): super(CNNTransformerPap, self). init ()

[0077] # CNN backbone selfcnn = nn. Sequential) nn.Conv2d(3, 32, 3, padding=l), nn.BatchNorm2d(32), nn.ReLUQ, nn.MaxPool2d(2), nn.Conv2d(32, 64, 3, padding=l), nn.BatchNorm2d(64), nn.ReLUQ, nn.MaxPool2d(2), nn.Conv2d(64, 128, 3, padding=l), nn.BatchNorm2d(128), nn.ReLUQ, nn.MaxPool2d(2)

[0078] )

[0079] # Flatten and linear projection selfproj = nn.Linear(128 * 32 * 32, embed_dim)

[0080] # Transformer encoder encoder layer = nn. TransformerEncoderLayer( d_model=embed_dim, nhead=num_heads, dim_feedforward=512, batch_first=True

[0081] ) self. transformer = nn.TransformerEncoder(encoder_layer, num_layers=depth)

[0082] # Classification head selffc = nn. Sequential nn.Linear(embed_dim, 128), nn.ReLUQ, nn.Dropout(0.3), nn.Linear(128, num classes)

[0083] ) def forward(self, x): batch size = x.size(O) feats = self.cnn(x) feats = feats. view(batch_size, -1) emb = self.proj(feats).unsqueeze(l) trans out = self.transformer(emb) out = self.fc(trans_out[:, -1, :]) return out

[0084] EfficientNet + LSTM Model Pipeline

[0085] In addition to the CNN-Transformer model, the invention provides an EfficientNet + LSTM architecture optimized for sequential and multi-field cytology analysis.

[0086] The EfficientNet backbone extracts high-level feature maps from individual image patches. Features from sequential patches, representing contiguous slide regions, are fed into a bidirectional LSTM that models spatial progression as a temporal sequence.

[0087] This configuration is particularly suited for slides with sparse atypical cells distributed across large regions, allowing the model to integrate information over multiple fields of view and generate slide-level predictions.

[0088] Py Torch Example — EfficientNet + LSTM import torch import torch, nn as nn import torchvision, models as models class EfficientNetLSTMPap(nn.Module): def init (self, num_classes=5, lstm_hidden=256, lstm_layers=2): super(EfficientNetLSTMPap, self). init ()

[0089] # EfficientNet backbone selfbackbone = models. efficientnet_bO(pretrained=True) self. backbone. classifier = nn.Identity() # remove default classifier

[0090] # LSTM for sequence modeling selflstm = nn.LSTM( input_size=1280, hidden_size=lstm_hidden, num_layers=lstm_layers, batch_first=True, bidir ectional=True

[0091] )

[0092] # Classification head selffc = nn. Sequential nn.Linear(lstm_hidden*2, 128), nn.ReLUQ, nn.Dropout(0.3), nn.Linear(128, num classes)

[0093] ) def forward(self, x_seq): batch_size, seq_len, c, h, w = x_seq.size() feats = [] for t in range(seq_len): feat = self.backbone(x_seq[:, t, :]) feats . append(feat) feats = torch. stack(f eats, dim=l)

[0094] Istm out, > = self.lstm(feats) out = self.fc(lstm_out[:, -1, :]) return out

[0095] Py Torch Training Loop Example import torch from torch. utils. data import DataLoader

[0096] # Assume dataset returns (sequence_tensor, label) train loader = DataLoader(train_dataset, batch_size=8, shuffle=True) val loader = DataLoader(val_dataset, batch_size=8, shuffle=False) model = EfficientNetLSTMPap(num_classes=5) optimizer = torch. optim.AdamW(model.parameters(), lr=le-4) criterion = nn.CrossEntropyLoss(weight=torch.tensor([1.0,2.0,2.0,3.0,4.0])) # Example weights for epoch in range(50): model. train() for x seq, y in train loader: optimizer. zero_grad( ) outputs = model(x seq) loss = criterion(outputs, y) loss.backwardQ optimizer. stepQ

[0097] # Validation model. eval() val_loss = 0 correct = 0 total = 0 with torch. no gradQ: for x seq, y in val loader: outputs = model(x seq) val_loss += criterion(outputs, y).item() preds = outputs. argmax(dim=l) correct += (preds == y).sum().item() total += y.size(O) val_acc = correct / total print(f'Epoch {epoch+1 }, Vai Loss: {val_loss:.4f}, Vai Acc: {val_acc:.4f}")

Claims

ClaimsWhat is claimed is:

1. A computer-implemented system for automated analysis of cytology slides, comprising: o an image acquisition module configured to receive digitized images of Pap smear slides; o a preprocessing module configured to normalize and segment said images into individual cellular regions; o a first artificial intelligence module comprising a convolutional neural network combined with a transformer encoder, configured to extract spatial and contextual features from said cellular regions and generate classification probabilities for multiple diagnostic categories; o a second artificial intelligence module comprising an EfficientNet backbone followed by a bidirectional long short-term memory network, configured to model sequential dependencies across contiguous image regions and generate classification probabilities; o a reporting module configured to aggregate outputs from said first and second artificial intelligence modules and produce a slide-level diagnostic prediction; o a data management and security module configured to store said images, metadata, and classification outputs with encryption, access control, and audit logging.

2. The system of claim 1, wherein the diagnostic categories comprise Negative for Intraepithelial Lesion or Malignancy (NILM), Atypical Squamous Cells of Undetermined Significance (ASC-US), Low-Grade Squamous Intraepithelial Lesion (LSIL), High-Grade Squamous Intraepithelial Lesion (HSIL), and Squamous Cell Carcinoma (SCC).

3. The system of claim 1 or 2, wherein the preprocessing module further comprises color normalization, noise reduction, and patch extraction of image regions.

4. The system of claim 3, wherein patch extraction is adaptive based on cellular density.

5. The system of claim 1, wherein the first artificial intelligence module produces interpretability outputs comprising attention heatmaps, saliency maps, and textual explanations.

6. The system of claim 1 , wherein the second artificial intelligence module integrates feature vectors from sequential patches corresponding to contiguous slide regions.

7. The system of claim 1, wherein the reporting module generates standardized diagnostic reports consistent with clinical guidelines.

8. The system of claim 1, wherein the data management and security module implements encryption of image data at rest and in transit, anonymization of patient identifiers, and role-based access control.

9. The system of claim 1 , further comprising an active learning subsystem that updates said artificial intelligence modules based on expert-reviewed corrections.

10. The system of claim 1, wherein said image acquisition module accepts images from both whole-slide scanners and conventional microscope cameras.

11. A method for automated analysis of cytology slides, comprising the steps of: acquiring digitized images of Pap smear slides; preprocessing said images to normalize, denoise, andsegment individual cells; applying a CNN-Transformer model to extract features and classify each image patch; applying an EfficientNet-LSTM model to extract sequential features and classify contiguous image regions; aggregating classification outputs to generate a slide-level prediction; producing a diagnostic report including confidence scores and interpretability outputs.

12. The method of claim 11, wherein said preprocessing further comprises stain normalization, geometric correction, and patch-based tiling.

13. The method of claim 11 or 12, wherein the CNN-Transformer model comprises multiple convolutional layers followed by multi-head self-attention layers.

14. The method of claim 11 or 12, wherein the EfficientNet-LSTM model comprises an EfficientNet backbone for feature extraction followed by a bidirectional LSTM network for sequential modeling.

15. The method of claim 11, wherein interpretability outputs include visual heatmaps indicating regions contributing most strongly to classification.

16. The method of claim 11 , further comprising a human-in-the-loop review in which slides with high abnormality probability are flagged for expert cytopathologist verification.

17. The method of claim 11, wherein the system continuously improves model performance via periodic retraining based on expert feedback and active learning.

18. The system of claim 1, wherein said system is deployable as on-premises software, private cloud, or public cloud service.

19. The system of claim 1, wherein said Al modules are containerized and support horizontal scaling with GPU acceleration.

20. The system of claim 1, further adapted for automated analysis of other cytology sample types, including urine, sputum, and fine-needle aspiration specimens.