Applying Deep Learning and Heatmap Techniques for Sinusitis Diagnosis Using Enhanced 3-D Image Segmentation

Sinusitis, characterized by inflammation of the sinus cavities, presents diagnostic challenges due to the subjective and time-consuming nature of manual computed tomography (CT) scan annotations. This study introduces a deep learning (DL) method combined with heatmap analysis to automate 3-D segmentation of sinusitis in CT images. The proposed approach reduces the dependency on fully segmented annotations, instead requiring only binary classification labels during training. A classification model identifies areas likely affected by inflammation, highlighted by gradient-weighted class activation mapping (Grad-CAM) to produce heatmaps, which are refined using a watershed algorithm for precise segmentation. This approach reduces annotation preparation time from 30 min to under 1 min while achieving a Dice similarity coefficient (DSC) of 90.04%. Built on the EfficientNet-B2 architecture, the model outperforms state-of-the-art methods like U-Net, U-Net++, and statistical-based approaches. Additionally, a novel instrumentation system integrated with CT scanners calculates inflammation areas in real time, streamlining diagnostic workflows. This method accelerates the process while improving the consistency and reliability of sinusitis diagnosis, potentially transforming clinical practices with advanced image processing techniques.