Рақамли технологияларнинг назарий ва амалий масалалари Volume 9 Issue 2 (2026) · pp. 143-149
Mammographic lesion detection based on BCA-YOLO and BI-RADS regression
Туракулов, Ш.Х.
Abstract
This paper proposes BCA-YOLO, a YOLO-based architecture enhanced with three novel components. Bilateral Cross-Attention (BCA) block, which computes cross-attention between the feature maps of the left breast and the horizontally mirrored contralateral right-breast feature maps at each feature scale. It then introduces an asymmetry-based residual update controlled by a learnable scalar coefficient α. Inter-View Consistency (IVC) module, which integrates craniocaudal (CC) and mediolateral oblique (MLO) features for each side using a learnable channel-wise gating mechanism. Ordinal BI-RADS regression head, which replaces conventional categorical cross-entropy with a cumulative link model employing monotonically parameterized thresholds. A reference implementation of the BCA-YOLO head in PyTorch adds approximately 3.1 million trainable parameters to the YOLOv8/v11 backbone. The BCA module is initialized with α = 0, allowing the model to exactly reproduce the baseline model for each mammographic view at initialization while preserving compatibility with pretrained backbone networks.
маммографияобнаружение объектовдвусторонняя асимметрияперекрёстное вниманиеординальная регрессияYOLOглубокое обучениеmammographyobject detectionbilateral asymmetry
Metadata source: the journal's OAI-PMH archive · Sindex does not store the full text; it links to the source.