Mechanosensitive TRPV4 immunohistochemistry improves deep learning-based classification of ductal carcinoma in situ beyond H&E morphology

Abstract

Ductal carcinoma in situ (DCIS) spans a biologic continuum from atypical ductal hyperplasia (ADH) to high-grade lesions with variable risk of progression to invasive ductal carcinoma (IDC), yet morphologic assessment by hematoxylin and eosin (H&E) remains diagnostically limited, particularly at the benign versus ADH/low-grade DCIS boundary. TRPV4, a mechanosensitive ion channel with pathology-dependent subcellular localization in DCIS, offers a biologically motivated immunohistochemical (IHC) marker that may refine classification beyond routine H&E assessment. We tested whether deep learning models trained on TRPV4 IHC outperform H&E-based models across the DCIS progression spectrum. We assembled a multi-institutional cohort of H&E and TRPV4 IHC whole-slide images from 108 patients, comprising an internal development cohort (n = 69), an external test cohort (n = 39), yielding 24,248 annotated tiles. Histopathological tiles from annotated regions were grouped into four ordered classes: normal/benign, ADH/low-grade DCIS, high-grade DCIS, and IDC. Xception and EfficientNet-B0 convolutional neural networks were trained with patient-level three-fold cross-validation on the development cohort and evaluated as ensembles on the external test cohort. On external patient-level testing, H&E ensembles achieved macro-F1 values of 0.43–0.44 and macro-AUC values of 0.73–0.80, whereas TRPV4 IHC ensembles improved performance to macro-F1 values of 0.68–0.72 and macro-AUC values of 0.91–0.92, corresponding to a 54.5–67.4% relative improvement in patient-level macro-F1. Patient-level per-class analyses showed the largest AUC gains with TRPV4 IHC versus H&E for ADH/low-grade DCIS (0.94–0.95 versus 0.61–0.70) and IDC (0.77–0.85 versus 0.61–0.69). Per-class analyses showed the largest gains with TRPV4 IHC versus H&E for ADH/low-grade DCIS (AUC, 0.83–0.84 versus 0.70–0.81) and IDC (AUC, 0.74–0.79 versus 0.65–0.66). These findings support TRPV4 IHC as a mechanistically grounded complement to H&E that improves patient-level discrimination across the DCIS progression spectrum, with the strongest gains for ADH/low-grade DCIS and IDC, in a pilot multi-institutional setting.