Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Atypical Mitosis Classification with DenseNet121: Stain-Aware Augmentation and Hybrid Loss for Domain Generalization

Adinath Dukre, Ankan Deria, Yutong Xie, Imran Razzak

TL;DR

This work tackles cross-domain generalization for atypical mitosis recognition in histopathology by combining a DenseNet-121 backbone with stain-aware augmentation (Macenko) and an imbalance-aware hybrid loss that blends weighted BCE and focal loss. The method demonstrates robust domain generalization across scanner and staining shifts, achieving a final MIDOG25 test-balanced accuracy of $0.850$ and ROC-AUC of $0.927$, with sensitivity $0.892$ and specificity $0.809$. Key contributions include a stain-aware preprocessing pipeline with 60% random cropping, an integrated loss function to address severe class imbalance, and extensive cross-domain evaluation showing improved discrimination over baselines. These results suggest a practical and scalable approach for robust atypical mitosis classification in real-world computational pathology workflows.

Abstract

Atypical mitotic figures are important biomarkers of tumor aggressiveness in histopathology, yet reliable recognition remains challenging due to severe class imbalance and variability across imaging domains. We present a DenseNet-121-based framework tailored for atypical mitosis classification in the MIDOG 2025 (Track 2) setting. Our method integrates stain-aware augmentation (Macenko), geometric and intensity transformations, and imbalance-aware learning via weighted sampling with a hybrid objective combining class-weighted binary cross-entropy and focal loss. Trained end-to-end with AdamW and evaluated across multiple independent domains, the model demonstrates strong generalization under scanner and staining shifts, achieving balanced accuracy 85.0%, AUROC 0.927, sensitivity 89.2%, and specificity 80.9% on the official test set. These results indicate that combining DenseNet-121 with stain-aware augmentation and imbalance-adaptive objectives yields a robust, domain-generalizable framework for atypical mitosis classification suitable for real-world computational pathology workflows.

Robust Atypical Mitosis Classification with DenseNet121: Stain-Aware Augmentation and Hybrid Loss for Domain Generalization

TL;DR

This work tackles cross-domain generalization for atypical mitosis recognition in histopathology by combining a DenseNet-121 backbone with stain-aware augmentation (Macenko) and an imbalance-aware hybrid loss that blends weighted BCE and focal loss. The method demonstrates robust domain generalization across scanner and staining shifts, achieving a final MIDOG25 test-balanced accuracy of and ROC-AUC of , with sensitivity and specificity . Key contributions include a stain-aware preprocessing pipeline with 60% random cropping, an integrated loss function to address severe class imbalance, and extensive cross-domain evaluation showing improved discrimination over baselines. These results suggest a practical and scalable approach for robust atypical mitosis classification in real-world computational pathology workflows.

Abstract

Atypical mitotic figures are important biomarkers of tumor aggressiveness in histopathology, yet reliable recognition remains challenging due to severe class imbalance and variability across imaging domains. We present a DenseNet-121-based framework tailored for atypical mitosis classification in the MIDOG 2025 (Track 2) setting. Our method integrates stain-aware augmentation (Macenko), geometric and intensity transformations, and imbalance-aware learning via weighted sampling with a hybrid objective combining class-weighted binary cross-entropy and focal loss. Trained end-to-end with AdamW and evaluated across multiple independent domains, the model demonstrates strong generalization under scanner and staining shifts, achieving balanced accuracy 85.0%, AUROC 0.927, sensitivity 89.2%, and specificity 80.9% on the official test set. These results indicate that combining DenseNet-121 with stain-aware augmentation and imbalance-adaptive objectives yields a robust, domain-generalizable framework for atypical mitosis classification suitable for real-world computational pathology workflows.
Paper Structure (21 sections, 3 equations, 1 figure, 3 tables)

This paper contains 21 sections, 3 equations, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Architecture Overview
  4. Dataset
  5. Preprocessing and Augmentation
  6. Network Architecture
  7. Imbalance-Aware Optimization
  8. Notation.
  9. Weighted BCE (WBCE).
  10. Focal loss.
  11. Combined objective.
  12. Implementation notes.
  13. Symbols and meanings.
  14. Training Protocol
  15. Evaluation Metrics
  16. ...and 6 more sections

Figures (1)

  • Figure 1: Overview of the proposed DenseNet121-based framework for atypical mitosis classification under domain variability. The pipeline takes mitotic figure patches from multiple domains, applies stain-aware preprocessing (Macenko normalization), and uses a DenseNet121 backbone followed by a binary classification head. The training strategy integrates 60% spatial cropping, stain perturbations, and a composite loss combining weighted cross-entropy, focal loss, and sampling-based imbalance handling. The model is evaluated using balanced accuracy, sensitivity, specificity, and ROC-AUC.