Computer-aided system and method for classification of lung nodules

Abstract

A non-invasive computer-aided system and method for classification of lung nodules includes receiving as input three distinct data representations describing a subject lung, extracting relevant features from each representation, analyzing the features using a transformer-based machine learning model, and classifying, using the model, nodules as benign or malignant.

Description

COMPUTER-AIDED SYSTEM AND METHOD FOR CLASSIFICATION OF LUNG NODULESCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of United States provisional patent application serial no. 63711383, filed October 24, 2024, incorporated herein by reference.FIELD OF THE INVENTION

[0002] A non-invasive computer-aided system and method for classification of lung nodules includes receiving as input three distinct data representations describing a subject lung, extracting relevant features from each representation, analyzing the features using a transformer-based machine learning model, and classifying, using the model, nodules as benign or malignant.BACKGROUND OF THE INVENTION

[0003] Lung cancer is the leading cause of cancer-related mortality globally, presenting a significant public health challenge. In 2025, it is estimated to be responsible for 124,730 deaths in the United States alone, accounting for 20.2% of all cancer deaths. With 226,650 new cases expected in the same year, lung cancer’s impact is not just numerical but deeply personal to patients and their families. The disease’s heterogeneity, with its diverse histological types, growth rates, and responses to treatments, demands a multifaceted approach to management, incorporating surgery, chemotherapy, radiotherapy, and targeted therapies. Early assessment of detected lung nodules is crucial for improving patient outcomes. However, it remains a formidable challenge due to the disease’s asymptomatic nature in its early stages. This difficulty often results in late diagnoses when treatment options are limited and less effective.

[0004] Recent advancements in artificial intelligence (Al), particularly through machine learning and deep learning, have revolutionized lung cancer screening by improving diagnostic accuracy and reducing false positives. Further innovations have explored the application of saliency-based capsule networks and efficient transfer learning to minimize computational demands while enhancing model generalization. The role of pre-trained models has emerged as pivotal, enabling the leveraging of extensive datasets for improved feature extraction without heavy preprocessing. However, current work in this area is limited by the gap between integrating local radiomics extracted by state-of-the-art deep-learning approaches and global radiomics extracted by state-of-the-art machine learning algorithms.This gap affects the accuracy of developing new approaches for the early assessment of detected lung nodules.SUMMARY

[0005] The present invention addresses these needs by introducing a novel representation of global radiomics that is rotationally, translationally, and scale invariant. This representation is presented alongside the raw computed tomography (CT) data, allowing for easy integration with original medical data in any deep-learning diagnostic approach. Central to this approach is the fusion of three distinct data representations: original Hounsfield Unitbased Volumes of Interest (VOIs); rotation, scale, and translation invariant Spherical Modified Local Ternary Pattern (SMLTP) capturing local textural descriptors; and spherical harmonics (SH) metrics detailing morphological characteristics derived from tangent isosurfaces. These representations are then processed through a convoluted neural network, such as, for example, Efficient-Net, for feature extraction. This amalgamation is innovatively analyzed using a transformer-based model, which treats the diverse feature sets as sequences of embeddings to capture both local and global patterns indicative of malignancy or benignity. Evaluated on a curated private dataset and the independent Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) benchmark, the machine learning model demonstrates exceptional performance, achieving unparalleled accuracy, sensitivity, and Area Under the Curve (AUC) metrics, thus underscoring its superior predictive capability and generalizability across clinical settings. Notably, the model’s 100% sensitivity rate in detecting true positives within the private dataset highlights its potential to significantly impact patient outcomes by reducing the likelihood of diagnostic oversights.

[0006] In one embodiment, the present invention is a computer-implemented method for classification of lung nodules comprising receiving medical image data of a subject lung; identifying and extracting at least one Hounsfield unit (HU)-based volume of interest (VOI) in the medical image data, the VOI including a representation of a lung nodule; generating, for each HU-based VOI, spherical modified local ternary patterns (SMLTP) and spherical harmonic (SH) metrics; extracting, using a machine learning model, at least one of textural features and morphological features from each HU-based VOI, each SMLTP, and SH metrics corresponding to the HU-based VOI; embedding the extracted features as vectors in the machine learning model; and classifying, using the machine learning model, the lung nodule as benign or malignant based at least in part on the extracted features. In another embodiment, the present invention is a non-transitory computer readable storage medium having computer program instructions stored thereon that, when executed by a processor, cause the processor to perform the following instructions: receive medical image data of asubject lung; identify and extract at least one Hounsfield unit (HU)-based volume of interest (VOI) in the medical image data, the VOI including a representation of a lung nodule; generate, for each HU-based VOI, spherical modified local ternary patterns (SMLTP) and spherical harmonic (SH) metrics; extract, using a machine learning model, at least one of textural features and morphological features from each HU-based VOI, each SMLTP, and SH metrics corresponding to the HU-based VOI; embed the extracted features as vectors in the machine learning model; and classify, using the machine learning model, the lung nodule as benign or malignant based at least in part on the extracted features. In some embodiments, the medical image data is computed tomography (CT) image data. In further embodiments, generating SMLTP includes, for each voxel in the VOI, defining a plurality of concentric spherical regions centered on the voxel, comparing an intensity of the voxel with voxels in each spherical region, classifying differences in intensity into one of a plurality of categories, and computing a new value for the voxel as a weighted sum of the categorizations. In certain embodiments, generating SH metrics includes partitioning the HU-based VOI into clusters based Hounsfield unit values, generating an iso-surface from each cluster, generating a 3D mesh from each iso-surface and representing each 3D mesh using a SH. In some embodiments, the machine learning model is a hybrid convoluted neural network (CNN) - transformer model. In further embodiments, the step of classifying the lung nodule as benign or malignant has at least one of an accuracy of at least 99% and a sensitivity of at least 99%. In certain embodiments, the step of classifying the lung nodule as benign or malignant has at least one of an accuracy of at least 99%, a sensitivity of at least 99%, and a specificity of at least 98%.

[0007] It will be appreciated that the various systems and methods described in this summary section, as well as elsewhere in this application, can be expressed as a large number of different combinations and subcombinations. All such useful, novel, and inventive combinations and subcombinations are contemplated herein, it being recognized that the explicit expression of each of these combinations is unnecessary.BRIEF DESCRIPTION OF THE DRAWINGS

[0008] A better understanding of the present invention will be had upon reference to the following description in conjunction with the accompanying drawings.

[0009] FIG. 1 is a schematic illustration of the disclosed transformer-based system for lung nodule classification.

[0010] FIG. 2 is a schematic illustration of the workflow of the SMLTP algorithm.

[0011] FIG. 3 is an illustration of the embedded local descriptors using the tangent isosurfaces reconstruction error quantified via spherical harmonics analysis (SH) and the non-homogeneity detected by spherical modified local ternary pattern (SMLTP), wherein high error and high non-homogeneity are indicated by darker color and low error and low nonhomogeneity are indicated by lighter color.

[0012] FIG. 4 is a schematic illustration of 3D mesh processing and spherical harmonic analysis.DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] The details of one or more embodiments of the presently-disclosed subject matter are set forth in this document. Modifications to embodiments described in this document, and other embodiments, will be evident to those of ordinary skill in the art after a study of the information provided in this document. The information provided in this document, and particularly the specific details of the described exemplary embodiments, is provided primarily for clearness of understanding and no unnecessary limitations are to be understood therefrom. In case of conflict, the specification of this document, including definitions, will control.

[0014] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the presently-disclosed subject matter belongs. Although any methods, devices, and materials similar or equivalent to those described herein can be used in the practice or testing of the presently-disclosed subject matter, representative methods, devices, and materials are now described.

[0015] Following long-standing patent law convention, the terms “a”, “an”, and “the” refer to “one or more” when used in this application, including the claims. Thus, for example, reference to “a cell” includes a plurality of such cells, and so forth.

[0016] Unless otherwise indicated, all numbers expressing quantities of ingredients, properties such as reaction conditions, and so forth used in the specification and claims are to be understood as being modified in all instances by the term “about.” Accordingly, unless indicated to the contrary, the numerical parameters set forth in this specification and claims are approximations that can vary depending upon the desired properties sought to be obtained by the presently-disclosed subject matter. As used herein, the term “about,” when referring to a value or to an amount is meant to encompass variations of ±10% of the most precise digit in the value or amount (e.g., “about 1” refers to 0.9 to 1.1, “about 1.1” refers to 1.09 to 1.11, etc.). The term “substantially,” when modifying a term associated with a number, has the same meaning as “about” (e.g., “substantially perpendicular” to an element means an orientation with ±10% of 90 degrees with respect to that element).

[0017] As used herein, ranges can be expressed as from “about” one particular value, and / or to “about” another particular value. It is also understood that there are a number of values disclosed herein, and that each value is also herein disclosed as “about” that particular value in addition to the value itself. For example, if the value “10” is disclosed, then “about 10” is also disclosed. It is also understood that each unit between two particular units are also disclosed. For example, if 10 and 15 are disclosed, then 11, 12, 13, and 14 are also disclosed.

[0018] This disclosure leverages a combined dataset of private curated CT scans and a publicly available LIDC-IDRI dataset to leverage robustness and generalizability of the proposed study. A total of 707 CT scans (230 benign and 477 malignant) forms a private dataset spanning over 5 years (2017-2022) of lung biopsy records. The included scans are progressively segmented by experienced radiologists using 3D Slicer software. This separation technique assures the appropriate delineation of the nodule, thus accurate training and validation of the model. Additionally, the LIDC-IDRI dataset was used for external validation. Including only nodules with a minimum size of 3 mm and consistent grading by radiologists, resulted in a dataset of 679 nodules (364 benign and 315 malignant).

[0019] Both datasets are preprocessed to optimize the data robustness and consistency for effective analysis. Initially, raw DICOM format CT images and binary segmentation maps are resampled to isotropic resolutions, matching the original voxel spacing for uniformity. The preprocessing steps involve resampling raw DICOM CT images and their associated binary segmentation maps to achieve isotropic voxel spacing. This is a standard process to ensure that the voxel dimensions (height, width, and depth) are consistent across all images.Resampling is done by interpolating the original data to match a uniform voxel spacing, typically achieved through trilinear interpolation. This process is for maintaining consistency in the spatial resolution of the images, which directly affects the accuracy of subsequent feature extraction methods, such as SMLTP and SH analysis. By ensuring uniform voxel sizes, discrepancies caused by varying resolutions in the original CT scans are eliminated, thereby optimizing the dataset for robust machine learning analysis and enhancing the data's robustness and consistency for effective nodule classification. Pulmonary nodule regions are then delineated using the binary maps to preserve the integrity of nodule boundaries. Further, cubic VOIs centered on the nodules are extracted based on centroids within the binary segmentations, ensuring complete encapsulation of the pulmonary nodule areas. These VOIs are then standardized to uniform dimensions through trilinear interpolation, achieving necessary scale consistency for subsequent analysis phases.

[0020] Compared to benign nodules, malignant nodules display rapid, disorganized growth patterns, resulting in distinctly irregular textures and morphologies. Based on this observation, and referring now to FIG. 1, the present invention comprises a novel system and method for classifying lung cancer nodules in CT images through combination of multiple methodologies to improve accuracy and interpretability as compared to known classification methods. The disclosed system and method involve the modification of raw CT images by preprocessing, followed by application of SMLTP and SH analysis. These techniques allow for the detailed extraction of textural and morphological features, which are then analyzed, along with the original Hounsfield unit-based VOIs, using transformer-based model to capture long-range dependencies and complex patterns within the data. The transformer-based model includes a hybrid model architecture that combined convoluted neural network (CNN) backbone with a transformer, incorporating positional encodings to impart spatial awareness. To address class imbalance and focus learning on challenging cases, the system employes the Focal Loss function, supplemented with L1 regularization to enhance model generalization. The system culminates in fusing three data representations, i.e., the HU-based VOIs, SMLTP, and reconstruction error-mapped iso-surfaces via SH, using transformer-based model adept to achieve superior classification accuracy. The original Hounsfield unit-based VOIs are identified and extracted using the segmentation technique described in U. S. Patent No. 8,731,255, but other techniques may be used in other embodiments.

[0021] Non-homogeneity Local Analysis using SMLTP’. This system and method extends the local binary pattern methodology through the introduction of the SMLTP algorithm, schematically summarized in FIG. 2. In SMLTP, for each voxel within the VOI, concentric spherical neighborhoods are defined, and a dual-threshold voxel comparison is performed to classify local intensity variations. The resulting ternary values are aggregated across spherical shells with radius-weighted encoding to generate a rotation-, scale-, and translation-invariant texture descriptor.

[0022] The essence of SMLTP lies in its dual-threshold strategy for evaluating voxel Hounsfield Unit (HU) variations, providing a refined mechanism to distinguish between benign and malignant nodules through comprehensive texture analysis. Key parameters, including thresholds (Tl, Th) for voxel difference calculations, the base (b) for rank calculation, and the maximum radius (rmax) for spherical region analysis, are specified to tailor the SMLTP computation to the unique characteristics of lung CT images.

[0023] For each voxel within the VOI, spherical regions with radii up to rmaxare considered. The centers and offsets of these spherical regions are computed, followed by generating spherical masks (SMr) that delineate each region’s boundaries as follows:(1)where c denotes the center of a voxel, and p represents points within the spherical region.

[0024] Each voxel’s value is updated based on SMLTP calculations within its concentric spherical neighborhood regions. This process involves comparing voxel intensity with neighboring voxels in spherical neighborhoods, classifying these intensity differences into three levels according to predefined thresholds, and computing a new voxel value as a weighted sum of these categorizations, adjusted by the base raised to the power of the sphere radius minus one, as shown in Equations 2 and 3:(2)(3)

[0025] Here, 1(c) and l(p) denote the intensity values of the central voxel and its neighbors, respectively, while Vnew(c) represents the updated voxel value.

[0026] Algorithm 1 outlines the step-by-step SMLTP processing method, implemented to enhance VOI analysis for lung cancer classification.Algorithm 1. Computation of Spherical Modified Local Ternary Pattern (SMLTP) 1: Input: Binary mask B, intensity volume I, thresholds (Ti, Th), base b, maximum radius rmax2: Output: SMLTP-encoded volume Vnew3: for all voxels c with B(c) 0 do4: Initialize v <— 05: for r = 1 to rmaxdo6: Compute shell SMraround c7: Initialize s <— 08: for all p e SMrdo9: Evaluate 0>(l(p), l(c), Ti, Th)10: Accumulate s <— s + <t>( )11: end for12: Update v <— v + s ■ br-113: end for14: Assign Vnew(c) v15: end for16: return Vnew

[0027] The sequential evaluation of texture patterns within concentric spherical neighborhoods of incremental radii offers a layered insight into the textural complexity of lung nodules, from macroscopic to microscopic details. The multi-scale perspective not only allows for the detection of subtle textural distinctions indicative of malignancy, as indicated in FIG. 3, but also achieves rotation, scale, and translation invariance, ensuring consistent analysis across nodules of varying sizes, shapes, and orientations within the 3D space. This invariance arises from the use of concentric spherical neighborhoods, which remove orientation dependence, and the radius-weight encoding, which normalizes scale and positional variations.

[0028] Morphology Local Analysis using SH: Morphological variability of pulmonary nodules was analyzed using SH, which provide a compact, rotation-invariant representation of 3D surfaces. In contrast to prior SH-based approaches that typically rely on a single smoothed surface, the disclosed method, schematically summarized in FIG. 4, includes three innovations. First, HU-guided tangent iso-surfaces are extracted, ensuring that the generated meshes reflect intensity-specific anatomical boundaries rather than purely geometric approximations. Second, each iso-surface is treated as an independent mesh, allowing localized morphological descriptors without over-smoothing or loss of detail. Third, SH reconstruction errors are explicitly mapped across iso-surfaces, providing a quantitative measure of local shape irregularities that may correlate with malignancy. Together, these steps yield a multi-scale, density-aware morphological signature that complements the SMLTP descriptors and HU-based volumetric data.

[0029] The VOI is initially segmented by adaptive clustering of voxel intensities, partitioning HU values into k discrete clusters, C = {Ci, C2,..., Ck}, which correspond to distinct HU bands. This clustering is achieved by minimizing intra-cluster variance via k-means:(4)where x represents voxel intensities and p; the centroid of cluster Cj. Each cluster corresponds to a specific HU band, which provides a clinically relevant segmentation of tissue densities within the nodule.

[0030] For each HU-defined cluster, an HU-constrained, tangency-enforced region-growing algorithm is applied, with region expansion bounded by both HU intensity limits and radial distance from the tumor’s centroid. This approach generates iso-surfaces, Qk, that conform to HU-defined boundaries while maintaining a uniform spatial alignment to the core,producing layers that are spatially and intensity-guided. Formally, each HU-guided isosurface Qk is defined as the set of voxels constrained by both HU intensity and a radial band: (5)Where l(x) denotes the HU intensity at voxel x, Ck is the HU cluster band, Co is the VOI centroid, and [r^,r ] is the radial band associated with Ck.

[0031] To further enhance tangency, each Qk undergoes 3D Gaussian smoothing, ensuring coherence and reducing high-frequency noise.

[0032] The smoothed iso-surfaces, now geometrically aligned and intensity-based, are converted into 3D meshes using the Marching Cubes algorithm. This is achieved by first applying a morphological closing operation to the generated iso-surfaces to improve voxel connectivity. Given an iso-surface Qk, we apply morphological closing to enhance structural integrity:(6)where C denotes the binary closing operator and B(1) is a ball structuring element with radius 1. The mesh is then generated using the Marching Cubes operator M:(7)(XF) = M((where X = {x "=1are the vertices and F the triangular faces.

[0033] The generated mesh undergoes a series of Laplacian flow smoothing iterations to reduce mesh noise while preserving essential geometric details. The Laplacian operator A is applied to each vertex Xj, updating its position as:(8)with(9)Where A is the smoothing factor, N(i) denotes the neighbors of Xj, and t indexes the iteration step. These HU-guided iso-surface meshes are treated as independent layers and normalized for SH analysis.

[0034] SH is applied to the smoothed 3D meshes for compact encoding and detailed morphological characterization. Each mesh is first converted from Cartesian coordinates{% "=1to spherical coordinates (n, 0j, <t>i) relative to the VOI centroid. SH basis functions y™(0, <[)) of degree I and order m are defined as:(10)where P™ denotes the associated Legendre polynomial. These basis function form an orthonormal set over the unit sphere and are used to project the radial function r(0, 4) of the mesh onto the SH basis. The surface is approximated by a truncated expansion:(11)With coefficients computed as:(12)ip) y{’” (, d)* sin 0 dddd-

[0035] The resulting coefficients encode the shape’s frequency components, allowing for efficient surface representation and analysis. This transformation provides a mathematically rigorous means to characterize and compare the geometry of iso-surfaces in a rotationinvariant and scale-normalized framework.

[0036] The quality of mesh representation via SH is quantified by reconstruction error, calculated as the difference between the original mesh and its reconstruction from a subset of harmonics:(13)Errorwhere xtdenotes the reconstructed vertex positions. This error metric serves as a direct indicator of the SH’s efficiency in preserving geometric detail, offering morphological descriptors for lung nodules and guiding the selection of an optimal number of harmonics for accurate, efficient mesh representation.

[0037] After analyzing each iso-surface, each voxel in the VOI is mapped to the reconstruction error associated with its corresponding iso-surface layer. The disclosed system and method quantifies iso-surface morphological precision using a predetermined number of harmonic steps, chosen empirically for optimal representation. By dissecting nodules into sequences of iso-surfaces and subjecting each to SH analysis, each VOI voxelis represented by the reconstruction error of its corresponding iso-surface, adapting nodules’ morphological insights from global to local descriptors, as schematically displayed in FIG. 3. The figure illustrates the fusion of global morphological descriptors and local textural descriptors through the combination of SMLTP and SH analysis. The global descriptors provided by SH capture the overall shape and structure of the nodule, while the local descriptors provided by SMLTP capture fine-grained textural non-homogeneity at the voxel level. By using iso-surface reconstruction errors from SH as local descriptors, the system is able to transform global shape insights into localized representations for each voxel. This layered analysis provides a multi-scale perspective, allowing the system to capture both macro and micro-level irregularities in the nodule, which are useful for distinguishing between benign and malignant cases.

[0038] Algorithm 2 outlines the step-by-step SH processing method.Algorithm 2. 3D Mesh Processing and Spherical Harmonics Analysis Pipeline1: Input: VOI volume V, HU partition C = {Ck}k=1K, centroid c0, radial bands {[rk-,rk+]}, number of smoothing iterations T, Laplacian flow parameter λ, SH expansion degree L 2: Output: Voxel-wise reconstruction error map3: Step 1: Mesh Generation4: for k = 1 to K do5: Extract iso-surface Ok from Ck using HU + radial band constraints6: Apply morphological closing: Q'k = C(Qk; B(1))7: Generate mesh: (Xk, Fk) = M(Q'k), with Mk = (Xk, Fk)8: Step 2: Laplacian Mesh Smoothing9: for t = 1 to T do10: for all vertices xi∈ Xkdo11: Compute discrete Laplacian Δxifrom neighbors N(i) 12: Update vertex: xi← xi+ λΔxi13: end for14: end for15: Step 3: Spherization16: Center mesh at centroid c017: Normalize vertices onto unit sphere18: Step 4: Spherical Harmonics Projection19: Convert vertices to spherical coordinates (ri, θi, φi)20: Form SH basis Ylm(θi, φi) up to degree L21: Estimate SH coefficients {alm} for r(θ, φ)22: Step 5: Reconstruction and Evaluation23: Reconstruct surface r̂(θ, φ) from {alm}24: Map back to 3D vertices x̂i25: Compute normalized reconstruction error for layer k26: Step 6: Voxel-wise Mapping27: Assign Errorkto all voxels in Ckwithin [rk-,rk+]28: Update the VOI with per-voxel error values29: end for30: Return the updated VOI with voxel-level descriptors

[0039] Transformer-based Model’. The disclosed system and method utilizes a transformerbased model to comprehensively analyze and fuse disparate feature sets. This approach combines three distinct representations of input data: original HU-based VOIs, SMLTP-generated VOIs, and reconstruction error-mapped VOIs derived from SH analysis of tangent iso-surfaces. To extract meaning features from these input data, a custom 3D CNN was designed. The CNN serves as the backbone feature extractor for the fusion model. Each array type encapsulates unique aspects of lung nodule morphology and texture.Normalization is applied to each 3D array to standardize the data, resizing them to 128 x 128 x 128 voxels to ensure uniform shape for subsequent analysis. This architecture eschews conventional pretrained networks, opting instead for a custom configuration that is rigorously designed to handle the unique spatial and morphological characteristics of the three data representations. The network architecture leverages sequential layers of 3D convolutions, each coupled with batch normalization and ReLU activations, to process the volumetric data inherent in medical imaging and provide three feature vectors for each nodule.

[0040] The input to the model consists of three feature vectors, each of length 512: HU-based VOIs, SMLTP-generated VOIs, and reconstruction error-mapped VOIs derived from SH analysis of tangent iso-surfaces. These vectors are normalized to the range [0, 1] based on the minimum and maximum values observed in the training dataset. Each feature vector is treated as a distinct token, effectively transforming each subject’s data into a sequence of three tokens. This results in an input tensor of shape (batch size, 3, 512), to which multiple layers of self-attention and feedforward networks are applied. The multi-head self-attention mechanism, central to the transformer, computes dynamic weights, enabling the model to capture intricate interactions among features. Given an input sequence X ∈ RN×d, attention is calculated as follows:(14)where Q, K, and V ∈ ℝN x drepresent the query, key, and value matrices, respectively.

[0041] To incorporate sequence ordering, positional encodings are added to the input features:(15)These encodings enhance the model’s ability to utilize positional information effectively.

[0042] Each transformer layer also includes a feedforward network applied independently to each token, defined as:(16)FFN(x) = max(0, xW1+ b1)W2+ b2,which introduces non-linearity and regularization. Residual connections and layer normalization are employed to facilitate gradient flow and stabilize training. After processing through multiple layers, the transformer generates a set of contextual embeddings, which are subsequently fed into a binary classification layer for the final diagnostic decision.

[0043] The overall architecture of the model can be expressed as follows:(17)zo= x + PE(18)Zlattn= LayerNorm(MultiHeadAttention(Zl-1) + Zl-1)(19)(20)Y = Linear(ZL)where X represents the input, PE denotes positional encoding, L is the total number of layers, and Y is the model’s output. Additionally, Zo is the input with positional encoding added, Z “ndenotes the output after the multi-head attention mechanism and residual connection in layer I, and Zi represents the output after the feedforward network and residual connection in layer I.

[0044] To address class imbalance and enhance robustness to difficult samples, we employ the Focal Loss(21)FL(pt) = -αt(1 - pt)γlog(pt),Where ptis the predicted probability of the true class, atis a class-balancing weight, and y adjusts the focus on hard examples. To further improve generalization, we add an L1 regularization term, yielding the overall objective:(22)where β controls the penalty strength and encourages sparsity in the model parameters.

[0045] The transformer-based model can synthesize different insights through integrating diverse data representations into a cohesive diagnostic framework. The disclosed system and method integrates three distinct feature sets - HU-based VOIs, SMLTP-transformed VOIs, and SH-transformed VOIs - into a unified transformer-based machine learning system. This integration allows the system to handle rotation-, translation-, and scale-invariant features, ensuring that it can accurately classify nodules regardless of their orientation or size in the CT scan. The system’s strength comes from its ability to extract and fuse both local descriptors (from SMLTP and SH) and global descriptors (from HU-based VOIs), which are processed through a convoluted neural network for feature extraction. The transformer model then synthesizes these features, learning the relationships between them to make highly accurate classifications. Moreover, while this system is discussed with respect to lung nodule classification, it provides a flexible and scalable architecture that can be adapted to other domains of medical imaging. The combination of local descriptors and transformerbased fusion offers a blueprint for tackling various diagnostic challenges across oncology and beyond.

[0046] Hyperparameter Tuning-. Components of the disclosed system and method, including SMLTP, SH, and the hybrid CNN-Transformer model are preferably tuned to obtain optimal performance. To identify the best configurations, both grid search and Bayesian optimization techniques were employed, systematically exploring the hyperparameter space. The final hyperparameter settings for each component of the model are summarized in Table 1, optimized through extensive tuning across multiple experiments to ensure peak performance on the lung nodule classification task.Table 1: Hyperparameters, search ranges, and final values

[0047] The performance of the disclosed system and method is evaluated in terms of accuracy, sensitivity, specificity, F1-score, and the area under the receiver-operating characteristics curve (AUC). Testing was conducted on two different datasets: a private dataset and the benchmark LIDC-IDRI dataset. For internal validation, an 80 / 20% trainingtesting split was applied to the private dataset. After training the model on the private data set, it was subsequently tested on the LIDC-IDRI dataset to assess the model’s effectiveness and general applicability.

[0048] First, the disclosed model was assessed utilizing the fusion of the three distinct local descriptors: the original HU-based VOIs, the SMLTP descriptors, and the reconstruction error-mapped SH, as derived from the private dataset. After 100 training epochs on the private dataset, the trained model was tested on the unseen LIDC-LDRI dataset. With less than 0.5% difference between both evaluations, the results highlight the robustness and generalizability of the diagnostic model. To analyze each feature’s contribution to the fusion model, individual tests were conducted on each descriptor as well as pairwise fusions (i.e., HU-based with SMLTP, HU-based with SH, and SMLTP with SH).

[0049] As shown Table 2, the disclosed fusion-based approach marked a significant improvement in performance across both datasets, highlighting its ability to efficiently process a variety of inputs with great effectiveness. On the private dataset, the fusion model achieved the best performance, reaching 99.30% accuracy, 100.0% sensitivity, 97.83% specificity, an F1-score of 99.48%, and an area under the curve (AUC) of 1.0. While single descriptors such as SMLTP (97.89% accuracy, AUC 0.9986) and SH (97.18% accuracy, AUC 0.9916) performed well, their pairwise combinations yielded further improvements. The SMLTP and SH fusion, for instance, raised specificity to 95.65% and improved the F1 -score to 98.97%, while the HU-based and SMLTP combination achieved 98.59% accuracy with an AUC of 0.9995. The HU-based and SH pairing also strengthened performance, reaching 97.89% accuracy and a 98.46% F1-score.Table 2: Experimental results of the disclosed lung nodules diagnostic model using different input feature sets on both private and benchmark (LIDC-IDRI) datasets

[0050] Comparable patterns emerged on the benchmark dataset. SMLTP alone achieved 97.50% accuracy and an AUC of 0.9977, surpassing HU-based features, but additional gains were observed when features were combined. The SMLTP and SH fusion attained 97.94% accuracy, a 97.79% F1-score, and the highest AUC among the pairwise models (0.9982). The HU-based and SMLTP pairing also performed strongly, with 97.64% accuracy and 97.48% F1-score, while the HU-based and SH combination delivered 95.73% accuracy and a 95.52% F1-score. The full fusion model consistently outperformed all alternatives, achieving 99.12% accuracy, 99.68% sensitivity, 98.63% specificity, a 99.05% F1-score, and an AUC of 0.9993.

[0051] The improvement in sensitivity ensures that malignant nodules are correctly identified, reducing the likelihood of missed diagnoses, which is critical in life-threatening conditions like lung cancer. Simultaneously, the increase in specificity minimizes false positives, ensuring that patients are not subjected to unnecessary follow-up procedures, reducing both healthcare costs and patient anxiety. These results indicate the disclosed model’s ability to significantly boost both the precision and reliability of its predictions.Overall, these findings underscore the value of integrating heterogeneous descriptors. HU-based features capture radio-density patterns, SMLTP encodes subtle textural inhomogeneities, and SH reconstruction errors reflect geometric irregularities. Their integration leverages these complementary perspectives to reduce both false positives and false negatives as compared to predictive models based on single descriptors.

[0052] The disclosed system and method was evaluated against several state-of-the-art architectures: Vision Transformer (ViT), EfficientNet, and a hybrid CNN-XGBoost approach. These models were trained and tested on raw CT scans, allowing them to exploit their capacity for large-scale spatial feature extraction, whereas the disclosed model operated on delineated nodule regions enriched with fused descriptors. This setup highlights the distinctive strengths of each paradigm: while conventional deep-learning models leverageglobal image context, the approach disclosed herein integrates localized intensity, textural, and geometric descriptors. As summarized in Table III, the proposed framework consistently surpassed all competing methods on both datasets.Table 3: Comparison of models

[0053] On the private dataset, the disclosed model achieved an accuracy of 99.30%, surpassing ViT, EfficientNet, and CNN-XGBoost models, which achieved 95.77%, 93.66%, and 95.77%, respectively. The model’s sensitivity of 100.0% indicates an unparalleled ability to correctly identify malignant nodules, outperforming the other architectures on this metric. Whereas the specificity, on the other hand, was 97.83% for the fusion-based model, enhancing robustness by reducing false positives. Additionally, the F1-score of 99.48% reflects a strong balance between precision and recall, while an AUC of 1.00 supports diagnostic reliability.

[0054] The results on the benchmark dataset further substantiate the robustness of this model and its capability for generalization. The model’s accuracy was significantly higher at 99.12%, compared to ViT at 93.37%, EfficientNet at 92.19%, and CNN-XGBoost at 94.11%. A sensitivity of 99.68% and a specificity of 98.63% show the model’s capability to maintain diagnostic performance on unseen data with minimal degradation. This model achieved an AUC of 0.9993 on the benchmark dataset, demonstrating high discrimination between benign and malignant nodules.

[0055] Moreover, the disclosed model achieves substantial improvements over previously reported models. Compared to models reported in Shafi et al. (Shafi et al., “An effective method for lung cancer diagnosis from CT scan using deep learning-based support vector network,” Cancers, vol. 14, no. 21, pp. 5457, 2022), Lima et al. (Lima et al., “Automatic classification of pulmonary nodules in computed tomography images using pre-trained networks and bag of features,” Multimedia Tools and Applications, vol. 82, no. 27, pp.42977-42993, 2023), Gupta et al. (Gupta et al., “Udct: Lung cancer detection andclassification using u-net and darts for medical CT images,” Multimedia Tools and Applications, pp. 1-21, 2024), Gautam et al. (Gautam et al., “Lung cancer detection from thoracic CT scans using an ensemble of deep learning models,” Neural Computing and Applications, vol. 36, no. 5, pp. 2459-2477, 2024), and Zheng et al. (Zheng et al., “Fine-grained contrastive learning for pulmonary nodule classification,” in 2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1-8), the disclosed model achieves the highest reported accuracy (99.12%), sensitivity (99.68%), and F1-score (99.05%) among comparable studies, while sustaining near-perfect AUC values. Overall, the disclosed model provides a clear improvement to previous technology-based solutions to the problem of lung cancer diagnosis by providing improved accuracy, sensitivity, F1 -score and AUC values as compared to earlier work in this field.

[0056] The foregoing detailed description is given primarily for clearness of understanding and no unnecessary limitations are to be understood therefrom. Modifications can be made by those skilled in the art upon reading this disclosure and may be made without departing from the spirit of the invention.

Claims

CLAIMS1. A computer-implemented method for classification of lung nodules, the method comprising:receiving medical image data of a subject lung;identifying and extracting at least one Hounsfield unit (HU)-based volume of interest (VOI) in the medical image data, the VOI including a representation of a lung nodule;generating, for each HU-based VOI, spherical modified local ternary patterns (SMLTP) and spherical harmonic (SH) metrics;extracting, using a machine learning model, at least one of textural features and morphological features from each HU-based VOI, each SMLTP, and SH metrics corresponding to the HU-based VOI;embedding the extracted features as vectors in the machine learning model; and classifying, using the machine learning model, the lung nodule as benign or malignant based at least in part on the extracted features.

2. The computer-implemented method of claim 1, wherein the medical image data is computed tomography (CT) image data.

3. The computer-implemented method of claim 1, wherein generating SMLTP includes, for each voxel in the VOI, defining a plurality of concentric spherical regions centered on the voxel, comparing an intensity of the voxel with voxels in each spherical region, classifying differences in intensity into one of a plurality of categories, and computing a new value for the voxel as a weighted sum of the categorizations.

4. The computer-implemented method of claim 1, wherein generating SH metrics includes partitioning the HU-based VOI into clusters based Hounsfield unit values, generating an iso-surface from each cluster, generating a 3D mesh from each iso-surface and representing each 3D mesh using a SH.

5. The computer-implemented method of claim 1, wherein the machine learning model is a hybrid convoluted neural network (CNN) - transformer model.

6. The computer-implemented method of claim 1, wherein the step of classifying the lung nodule as benign or malignant has an accuracy of at least 99% and a sensitivity of at least 99%.

7. A non-transitory computer readable storage medium having computer program instructions stored thereon that, when executed by a processor, cause the processor to perform the following instructions:receive medical image data of a subject lung;identify and extract at least one Hounsfield unit (HU)-based volume of interest (VOI) in the medical image data, the VOI including a representation of a lung nodule;generate, for each HU-based VOI, spherical modified local ternary patterns (SMLTP) and spherical harmonic (SH) metrics;extract, using a machine learning model, at least one of textural features and morphological features from each HU-based VOI, each SMLTP, and SH metrics corresponding to the HU-based VOI;embed the extracted features as vectors in the machine learning model; and classify, using the machine learning model, the lung nodule as benign or malignant based at least in part on the extracted features.

8. The non-transitory computer readable storage medium of claim 7, wherein the medical image data is computed tomography (CT) image data.

9. The non-transitory computer readable storage medium of claim 7, wherein generating SMLTP includes, for each voxel in the VOI, defining a plurality of concentric spherical regions centered on the voxel, comparing an intensity of the voxel with voxels in each spherical region, classifying differences in intensity into one of a plurality of categories, and computing a new value for the voxel as a weighted sum of the categorizations.

10. The non-transitory computer readable storage medium of claim 7, wherein generating SH metrics includes partitioning the HU-based VOI into clusters based Hounsfield unit values, generating an iso-surface from each cluster, generating a 3D mesh from each isosurface and representing each 3D mesh using a SH.

11. The non-transitory computer readable storage medium of claim 7, wherein the machine learning model is a hybrid convoluted neural network (CNN) - transformer model.

12. The non-transitory computer readable storage medium of claim 7, wherein the step of classifying the lung nodule as benign or malignant has an accuracy of at least 99% and a sensitivity of at least 99%.

Images