Human-AI Collaborative INSIGHT Diagnostic Workflow for in Breast Cancer With Extensive Intraductal Component

Overview

The goal of this clinical trial is to see if an artificial intelligence (AI)-assisted method helps doctors more accurately detect invasive breast cancer in people with a specific type of tumor called "extensive intraductal carcinoma" (EIC). This type of tumor is challenging to diagnose correctly using standard methods. The main question this study aims to answer is: Does the new AI-assisted method find more invasive breast cancer in EIC tumors compared to the standard method? Researchers will compare two groups: * Group 1 (INSIGHT): Doctors review breast tissue samples using an AI tool that highlights suspicious areas needing closer attention. * Group 2 (Conventional): Doctors review breast tissue samples without AI help, using the standard method. This comparison will show if the AI-assisted method works better at finding invasive cancer. What happens in the study? * Researchers will use stored breast tissue samples already collected during the participant's surgery. * Each sample will be randomly assigned to be reviewed using either the new AI-assisted method (Group 1) or the standard method (Group 2). * In Group 1, an AI program will scan the tissue images first and point out areas that might contain invasive cancer for the doctor to check closely. * In Group 2, doctors will review the tissue images without any AI help, using their standard process. * Researchers will measure which method finds invasive cancer more accurately, how long the review takes, and how many additional tests (called IHC stains) are needed. No new procedures are required from participants; the study uses existing tissue samples.

Breast cancer with extensive intraductal component (EIC) presents significant diagnostic challenges, characterized by widespread ductal carcinoma in situ (DCIS) frequently accompanied by small invasive foci (≤10 mm). Accurate identification of invasive carcinoma in EIC is critical for clinical staging and treatment decisions, yet conventional diagnostic methods face substantial limitations. Pathologists must manually screen extensive DCIS regions for minute invasive components, a labor-intensive process with reported miss rates reaching 20%, particularly for microinvasive foci (≤1 mm). Diagnostic uncertainty frequently leads to excessive immunohistochemical (IHC) staining (e.g., p63, CK5/6), with each stain costing ¥373.40, significantly increasing healthcare costs and prolonging turnaround times. To address these challenges, we developed the INSIGHT (INvasion Screening with Intelligent Guidance for Histopathology Triage) human-AI collaborative workflow. This solution integrates four public datasets (TiGER, BRACS, BACH, CAS\_PUIH) and employs weakly supervised pseudo-labeling to expand annotated pixels 22-fold to 25 billion, specifically improving representation of DCIS (3.14% to 12.53%) and benign tissue (0.65% to 10.9%). The AI model, based on a UperNet-VAN architecture, achieved Dice scores of 0.877 (training), 0.853 (validation), and 0.847 (testing). The system processes segmented invasive regions through size filtering (\>500 µm²) and cluster grouping to generate actionable regions of interest (ROIs) for pathologist guidance. In our preliminary retrospective study (576 whole slide images from 44 EIC patients), the INSIGHT workflow demonstrated superior diagnostic performance compared to conventional methods: sensitivity improved from 82.7% to 95.1% (p\<0.001), with particularly notable gains in detecting ≤1 mm microinvasive foci (69.4% to 91.8%); negative predictive value (NPV) reached 96.7% versus 89.6% (p\<0.001). The workflow reduced mean diagnostic time by 41.4% (102.6 to 60.1 seconds per slide, p\<0.001) and decreased IHC usage by 40.4% (p=0.011). While standalone AI showed high sensitivity (95.6%), its specificity remained limited (76.6%), underscoring the necessity of human-AI collaboration. This prospective clinical trial aims to validate the INSIGHT workflow's generalizability in real-world clinical settings, quantify its impact on patient stratification and treatment decisions, and establish standardized protocols for AI-assisted diagnosis to bridge critical gaps in computational pathology translation from research to clinical practice.