Automatic Bone Metastasis Diagnosis Framework via Lightweight Transformer-Based Network

Bone metastasis is one of the main cancer complications whose diagnosis is mainly based on bone scintigraphy. As an assistance for the diagnosis of bone metastasis, researches based on convolutional neural networks have achieved impressive performance. However, due to the absence of relevant dataset, most existing methods could only diagnose whether metastasis occurred without indicating the general location of its occurrence or make diagnoses of only few specific organs. Such models may lose the insights of regional information and deviate from clinical practice. To address this issue, we construct the first bone scintigraphy dataset where whole body images in the anterior view are segmented into 13 distinct body parts and all subsets of body parts are carefully annotated according to formal diagnoses. With comprehensively annotated dataset, we propose an automatic diagnosis framework via lightweight transformer-based network with early convolutions and spatial dimension reduction to make detailed diagnoses for all body parts. The results reveal that our overall diagnosis achieves higher recall, satisfying accuracy and with fewer parameters compared to popular models for medical image classification. Dataset and code would be made public upon acceptance.