Publication Details
Abstract
Computer-aided detection (CAD) of breast cancer in mammography has emerged as one of the most extensively studied tasks in artificial intelligence for medical imaging. However, practical deployment in hospitals located in low- and middle-income countries is constrained by three acute challenges: (i) the absence of bounding-box annotations in local DICOM images, (ii) clinical reports written simultaneously in three scripts — Uzbek Cyrillic, Uzbek Latin, and Russian — frequently with code-switching within a single document, and (iii) the lack of workflow integration that hinders the conversion of model outputs, after radiologist verification, into gold-standard labels for training. This article presents a ten-stage, fully reproducible pipeline that transforms raw clinical DICOMs and free-text radiology reports into a radiologist-verified, gold-labeled detection dataset. The system integrates four products: the MAMOGRAF annotation platform, the XS-Classifier multilingual classifier, the TILLNet-Det multimodal detector, and a multilingual weakly supervised training pipeline. A three-stage training curriculum exploits data of varying quality in a structured manner. Architectural and training contributions are isolated through controlled ablation studies. The full multimodal model has 34.53 million parameters, with the additional 2.42 million parameters of the text branch contributing the multimodal capability. The proposed zero-initialized FiLM fusion mechanism mathematically guarantees that the multimodal model is bit-equivalent to its image-only counterpart at initialization, eliminating cold-start instability and providing graceful degradation when reports are absent or incorrect.