Classification of pure DCIS cases in breast ultrasound images by multiscale contrastive learning

Distinction between invasive breast cancers and non-invasive cancers is important for determination of treatment planning. The decision is generally made based on the biopsy result; however, if diagnosis can be predicted by imaging, it may be useful for timely treatment planning and proper tissue sampling during biopsy. The purpose of this study is to classify breast ultrasound images with invasive cancers and non-invasive cancers. The number of cases used in this study is 690 breast ultrasound images, including 584 invasive cancers and 106 ductal carcinomas in situ (DCIS). Several image patches with different sizes were sampled including a lesion and at margin of a lesion to increase training samples. Since cases are highly imbalanced, the classification model was first pretrained for matched pairs and unmatched pairs with contrastive loss. The model is then fine-tuned for classification of invasive cancers and DCISs. ResNet was used as a base model. Models were trained with patches with different image sizes, and the results were ensembled. Although accuracy was slightly decreased from the baseline model, the recall rate for DCIS cases was greatly improved with an improvement in F1 score. Classification of non-invasive cancers on ultrasound images may support prompt treatment planning and biopsy procedures.