BSc-DARTS: Neural Architecture Search for Automatic Diagnosis of Bone Metastases in Bone Scintigrams

Bone scintigram is a highly effective medical imaging technique widely used for the rapid screening of bone metastases, contributing significantly to early disease detection and prognosis assessment. Neural architecture search enables automated design and optimization of network structures by leveraging data characteristics and task-specific requirements.

For the automated and accurate diagnosis of bone metastasis in bone scintigrams, this study proposes an improved differentiable neural architecture search framework to address two key limitations of the original DARTS method: (1) the insufficient representation capability of standard candidate operations for characterizing bone metastasis lesions, and (2) architecture degeneration.

Based on the DARTS framework, two novel candidate operations were developed, and the training supernet architecture was optimized. (1) The channel-attention-integrated residual operation incorporates synergistic channel-wise intelligent weighting and gradient stabilization mechanisms, providing an efficient yet highly discriminative feature transformation module for architecture search. (2) The spatial-attention-enhanced multibranch operation combines multi-scale feature fusion with spatially adaptive focusing, significantly improving the model’s ability to localize critical regions and detect lesions of varying sizes. (3) A dual-path convolutional structure is introduced into the training supernet to dynamically optimize both channel-wise and spatial dimensions of shallow features, thereby generating highly discriminative representations for subsequent cell architecture search.

Comprehensive evaluations on clinical datasets demonstrate the effectiveness of the proposed method, achieving an accuracy of 0.8451, precision of 0.8700, recall of 0.8447, F1-score of 0.8423, and an AUC of 0.92.

The results validate that incorporating manual network design experience into the architecture search process can significantly improve model performance. Systematically leveraging anterior-posterior views of the same case can improve lesion detection in complex anatomical regions. Future work will focus on developing a dual-view architecture search framework based on complementary feature fusion.

The proposed neural architecture search method effectively detects bone metastasis in bone scintigrams and outperforms existing state-of-the-art approaches. The newly developed candidate operations and supernet optimizations successfully address the limited representational capacity of standard convolutions in the original DARTS operations, leading to substantial improvements in classification performance.