Skip to content

Automated Detection of Malignant Lesions in the Ovary Using Deep Learning Models and XAI

Conference: CVPR 2026 arXiv: 2603.11818 Code: None Area: Medical Image Classification / Explainable AI Keywords: Ovarian Cancer Detection, CNN Comparison, Explainable AI, Histopathology, InceptionV3

TL;DR

This work systematically compares 15 CNN variants (LeNet/ResNet/VGG/Inception) on five-class classification of ovarian cancer histopathology images. InceptionV3-A (ReLU) is selected as the final model, achieving 94% across comprehensive metrics, with comparative explainability analysis conducted using three XAI methods: LIME, SHAP, and Integrated Gradients.

Background & Motivation

Background: Ovarian cancer is the 7th most common cancer among women worldwide and carries an extremely high mortality rate. A core challenge is the lack of effective early screening methods — unlike breast cancer (mammography) or cervical cancer (Pap smear), ovarian cancer can only be confirmed through invasive biopsy. Deep learning has shown progress in detecting various cancers, yet DL-based approaches for ovarian cancer remain limited.

Limitations of Prior Work: (1) Existing non-invasive detection methods (transvaginal ultrasound, CA-125 blood test, pelvic examination) lack sufficient accuracy to serve as reliable screening tools; (2) Definitive diagnosis relies on biopsy, which is invasive and time-consuming; (3) Existing DL approaches mostly employ single models, lacking systematic multi-architecture comparison and XAI-based explainability support, leading to low clinical trust and adoption.

Key Challenge: An automated detection system must simultaneously achieve high accuracy and interpretability; however, available histopathology datasets are extremely small (only 498 images), and high-performing models (VGG with transfer learning) are difficult to explain effectively via XAI due to frozen pretrained feature layers.

Goal: To build a high-accuracy classification model on a small-scale ovarian cancer histopathology dataset and provide transparent support for clinical decision-making through XAI.

Key Insight: Cast a wide net by comparing 15 CNN variants, selecting models based on both accuracy and XAI feasibility rather than solely pursuing peak performance.

Core Idea: Model selection should consider interpretability alongside accuracy — InceptionV3 (trained from scratch) is preferred over the higher-accuracy VGG (transfer learning) because the former is more amenable to effective XAI explanation.

Method

Overall Architecture

Acquire 5-class histopathology images (Clear Cell, Endometrioid, Mucinous, Non-Cancerous, Serous) totaling 498 images from the Mendeley dataset → Apply Albumentations-based data augmentation to expand to 2,490 images → Systematically train 15 CNN variants → Select the best model (InceptionV3-A) through comprehensive evaluation → Apply three XAI methods (LIME/SHAP/IG) for comparative explainability analysis.

Key Designs

  1. Data Augmentation Pipeline
  2. Albumentations library is used for rotations (up to 180°), horizontal/vertical flips, and random brightness/contrast/saturation/hue transformations.
  3. Each original image generates 4 augmented copies, expanding the dataset from 498 to 2,490 images (~498 per class, maintaining class balance).
  4. After conversion to tensors, RGB values are normalized from 0–255 to 0–1, significantly improving training stability.

  5. Random 80:20 train-test split (1,992 training / 498 testing).

  6. Systematic Comparison of 15 CNN Variants

  7. LeNet series (3 variants): Base (lr=0.001) / +Dropout / +Step Decay, trained for 100 epochs.
  8. ResNet series (4 variants): ResNet-34 at two resolutions (32×32 and 224×224), ResNet-50, and ResNet-101; optimal lr and dropout rate determined via random search (10 iterations × 3 epochs).
  9. VGG series (4 variants): VGG16-A/B/C and VGG19, all using ImageNet transfer learning with frozen feature layers and only fully connected layers trained.

  10. Inception series (4 variants): InceptionV1-A (ReLU) / V1-B (Tanh) / V3-A (ReLU + BatchNorm) / V3-B (Tanh + BatchNorm), all trained from scratch for 80 epochs.

  11. Comparative XAI Analysis Using Three Methods

  12. LIME: Locally interpretable; generates superpixel-level explanation maps (limited to 10 most important features), revealing local rationale behind predictions.
  13. Integrated Gradients: Gradient attribution method that integrates gradients along the path from a baseline to the input, producing pixel-level importance maps.
  14. SHAP: Shapley value-based attribution method quantifying the marginal contribution of each pixel to the prediction.
  15. Highlighted regions produced by all three methods exhibit significant overlap, validating the consistency and reliability of black-box explanations.

Loss & Training

Softmax output layer with cross-entropy loss. LeNet variants are trained for 100 epochs; ResNet variants use random search to determine optimal hyperparameters (lr range: 0.0001–0.1, dropout range: 0.0–0.9); VGG variants use ImageNet pretrained weights with frozen convolutional layers and only fully connected layers trained; Inception variants are trained from scratch for 80 epochs. The VGG series is excluded from final selection — despite achieving the highest accuracy, the frozen feature layers from transfer learning impede effective XAI application.

Key Experimental Results

Main Results

Model Accuracy Precision Recall F1-Score
VGG19 (Transfer Learning) 97.19% 97.31% 97.19% 97.20%
VGG16-A (Transfer Learning) 96.99% 96.98% 96.99% 96.97%
InceptionV3-A (Selected) 94.58% 94.75% 94.58% 94.62%
InceptionV1-B 85.74% 86.26% 85.74% 85.42%
LeNet-A 61.85% 62.20% 61.85% 61.96%
ResNet-34 (224) 57.03% 59.39% 57.03% 57.70%
ResNet-50 34.14% 47.75% 34.14% 33.47%

Although the VGG series achieves the highest scores, it is excluded because the black-box feature layers from transfer learning render XAI-based explanation ineffective.

Ablation Study

Comparison Dimension Conclusion
VGG16-A vs. VGG16-O (Kasture et al.) Same dataset: 96.99% (ours) vs. 84.64% (theirs) on augmented data; 77.78% vs. 50% (+27.78 pp) on original data, attributed to tensor conversion and normalization.
InceptionV3-A (ReLU) vs. InceptionV3-B (Tanh) 94.58% vs. 82.13%; ReLU significantly outperforms Tanh by 12.45 pp.
ResNet-34 32×32 vs. 224×224 43.78% vs. 57.03%; input resolution has a substantial impact on ResNet performance.
Three XAI Methods LIME/SHAP/IG highlighted regions show consistent overlap, validating explanation reliability.

Key Findings

  • Model complexity does not guarantee performance: ResNet-50/101 perform extremely poorly (34–43%), likely due to insufficient hyperparameter search and training epochs.
  • Data preprocessing (tensor conversion + normalization) yields substantial gains on small datasets, exceeding the baseline work by 27 pp.
  • XAI feasibility is an important dimension of model selection — a 2.4 pp accuracy trade-off in exchange for interpretability is considered worthwhile.

Highlights & Insights

  • The systematic comparison of 15 models provides valuable reference for CNN selection in medical imaging.
  • Incorporating XAI feasibility into model selection — rather than solely maximizing accuracy — reflects deployment-oriented and clinical-trust-driven considerations.
  • The comparative XAI analysis demonstrates the complementarity and consistency of LIME, SHAP, and IG.
  • The data augmentation and normalization strategy on an extremely small dataset offers useful practical insights.

Limitations & Future Work

  • The dataset is extremely small (498 original → 2,490 augmented images); generalizability is questionable and no cross-validation is performed.
  • Only classical CNN architectures are explored; ViT, medical pretrained models (e.g., BiomedCLIP), or more modern architectures are not investigated.
  • Multi-center/multi-institution data validation and real-world clinical testing are absent.
  • The ResNet series performs extremely poorly (34–57%); the hyperparameter search strategy (only 10 iterations × 3 epochs) is likely severely inadequate.
  • Per-class ROC-AUC details (e.g., confusion between Clear Cell and Serous subtypes) are not reported.
  • Augmentation strategies are limited to simple geometric and color transforms; more advanced techniques (e.g., MixUp, CutMix, Mosaic) are not explored.
  • vs. Kasture et al. (VGG16-O): On the same dataset, this work's VGG16-A achieves 96.99% vs. 84.64% on augmented data, demonstrating the value of data preprocessing.
  • vs. Hsu et al.: Their work employs ensemble learning with ResNet-18/50/Xception for ultrasound-based ovarian cancer detection, achieving higher accuracy but relying on larger datasets.
  • vs. Wang et al.: Their approach uses DL for MRI-based benign/malignant ovarian differentiation, reaching 87% accuracy, but on a different imaging modality.
  • The augmentation strategy for medical small-data scenarios and the concept of "interpretability-oriented model selection" offer practical reference for similar work.

Rating

  • Novelty: ⭐⭐ All methodological components follow standard practices; no new architecture or technique is proposed.
  • Experimental Thoroughness: ⭐⭐ Dataset is too small; generalization validation, cross-validation, and statistical significance testing are absent.
  • Writing Quality: ⭐⭐⭐ Structure is clear, but some descriptions are redundant and equation numbering is not tightly integrated with the main text.
  • Value: ⭐⭐ Provides some reference as an introductory medical AI study, but insufficient innovation and experimental depth to support a top-venue publication.