CCI-MUL: A Cervical Colposcopic Image Multitask Learning Framework for Diagnosis-Related Region Detection and Lesion...

Cervical cancer is a common malignancy threatening women's health, and early identification of cervical precancerous lesions is important for screening, intervention, and prognosis. However, cervical colposcopic images often suffer from uneven illumination, complex tissue textures, blurred lesion boundaries, low contrast, high interclass similarity, and large intraclass variation, which make diagnosis-related region localization and lesion classification challenging. To address these issues, this study constructed a multicenter dataset containing 3706 images and proposed CCI-MUL, a multitask learning framework for cervical lesion analysis. The main methodological innovation of CCI-MUL lies in its unified detection-classification design, which jointly optimizes region localization and lesion classification within a shared feature representation. The framework adopts a shared feature extraction network and integrates lesion detection and classification branches. The detection branch localizes suspicious regions, while the classification branch fuses upsampled multiscale features to achieve lesion recognition. To enhance feature representation, a gradient-contrast coupled convolution (GCCConv) module was designed to jointly model gradient and local contrast information, strengthening boundary, texture, and low-contrast lesion cues. A local context guidance unit (LCGU) was further proposed to strengthen the interaction between local lesion features and contextual semantics. In addition, a hierarchical reparameterized feature enhancement (HRep) module was embedded into the neck stage to refine multiscale fused features through re-parameterized enhancement, improving feature fusion for localization and classification. Fivefold cross-validation experiments showed that CCI-MUL achieved precision, recall, mAP@0.5, and mAP@0.5:0.95 of 0.925 ± 0.011, 0.909 ± 0.014, 0.948 ± 0.003, and 0.618 ± 0.004 in detection and precision, recall, specificity, and accuracy of 0.944 ± 0.009, 0.907 ± 0.011, 0.958 ± 0.003, and 0.947 ± 0.003 in classification, respectively.