Multi-Orientation Local Texture Features for Guided Attention-Based Fusion in Lung Nodule Classification

Computerized tomography (CT) scan images are widely used in automatic lung cancer detection and classification. The nodules’ texture distribution throughout the CT volume can vary significantly, accurate identification consideration of discriminative information this greatly help classification process. Deep stacks recurrent convolutional operations cannot entirely represent such variations, especially size location nodules. To model complex pattern inter/intra dependencies slices each nodule, a multi-orientation-based guided-attention module (MOGAM) is proposed paper, which provides high flexibility concentrating on relevant extracted from different regions nodule non-local manner. Moreover, to provide with finer-grained volume, specifically-designed local feature descriptors (TFDs) multiple orientations. These TFDs not only textural across but also encode approximate within slice. extended experimentation has shown effectiveness combination these through guided attention mechanism. According results obtained standard LIDC-IDRI dataset, approach outperformed other counterparts terms accuracy AUC evaluation metrics. Also, detailed explainability analysis provided, demonstrating correct functioning attention-based fusion approach, required by medical experts.