What question did this study set out to answer?

To develop a CNN pipeline that enhances early prediction of pCR to NAC in breast cancer through improved feature selection and interpretability.

February 2, 2026Open Access

Enhancing early prediction of pathological complete response in breast cancer using attention-based convolutional neural networks in digital pathology

Puntos clave

To develop a CNN pipeline that enhances early prediction of pCR to NAC in breast cancer through improved feature selection and interpretability.
Retrospective analysis of 384,076 tiles from 122 stained WSIs
Segmentation of annotated WSIs into tiles
Filtering of noninformative regions using Mini-Batch C-Fuzzy K-Means
Analysis of tiles using a CNN with a Convolutional Block Attention Module
Model achieved 81.4% AUC and 81.3% accuracy in the investigational cohort
Validation cohort showed 80.9% AUC and 80.0% accuracy
External validation cohort yielded 76.2% AUC and 70.0% accuracy
Sensitivity and specificity varied across cohorts but demonstrated robustness across different resolutions.

Resumen

Objective To develop an attention-based convolutional neural network (CNN) pipeline for early prediction of pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in breast cancer, improving feature selection and interpretability in whole slide image (WSI) analysis. Methods A retrospective analysis was conducted on 384,076 tiles extracted from 122 Hematoxylin and Eosin-stained WSIs, divided among an investigational cohort (IC, 82 patients enrolled at IRCCS Istituto Tumori “Giovanni Paolo II”), a validation cohort (VC, 20 patients, same Institution), and an external validation cohort (EVC, 20 patients belonging the Yale trastuzumab response cohort public dataset). WSIs were first annotated and then automatically segmented into tiles. Noninformative regions were filtered using Mini-Batch C-Fuzzy K-Means. Remaining tiles were analyzed using a CNN with a Convolutional Block Attention Module, prioritizing both histological features and tiles critical for predicting pCR. Results The model achieved robust performance: 81.4% AUC, 81.3% accuracy, 80.0% specificity, and 83.3% sensitivity in IC; 80.9% AUC, 80.0% accuracy, 85.78% specificity, and 66.7% sensitivity in VC; and 76.2% AUC, 70.0% accuracy, 71.4% specificity, and 66.7% sensitivity in EVC. The EVC, consisting of WSIs at 20× magnification compared to the 40× magnification of IC and VC, demonstrated the model's robustness to varying resolutions. Conclusion This is an innovative pipeline that not only improves prediction but also enhances the clinical utility of digital pathology.

Leer artículo completoexternamente

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo