Use of Spectral-Enhanced Transformers for Automated Head and Neck Tumor Segmentation for Radiotherapy Planning

Summary: Modern Radiotherapy for Head and neck cancer (HNC) treatment requires precise delineation of both tumour and normal structures on CT scans. This segmentation process is crucial for radiotherapy planning, but it is time-consuming and challenging. HNC tumours exhibit immense variability in size, shape, and location, posing a significant bottleneck for even skilled radiation oncologists. Existing commercial algorithms can segment standard anatomical structures, but automated tumour delineation remains elusive because of size, location and shape variability. Moreover, this manual process demands extensive training and time, affecting treatment efficiency and potentially introducing inter-observer variability in dose prescription. With this background, we intended to develop auto-image segmentation for tumours precisely and accurately We propose a novel AI-powered approach utilizing a spectral-enhanced transformer model for automated HNC tumour segmentation. This model builds upon the promising foundation of the MAXFormer architecture, further enhanced by integrating a spectral attention mechanism within the decoder phase. This innovative layer leverages frequency-domain information, empowering the model to capture intricate local details within complex HNC CT scans, ultimately leading to robust and accurate tumour delineation. Our initial dataset comprises axial CT scans from 92 HNC patients, meticulously annotated with ground truth tumour and normal structure segmentations by a trained radiation oncologist. These scans undergo rigorous preprocessing for network training and analysis. The enhanced MAXFormer model, encompassing an encoder-decoder backbone with a refined fused connection module, leverages the spectral attention mechanism within the decoder architecture to refine tumour boundary detection. This methodology builds upon our previous successes in automating normal organ segmentation using similar transformer-based approaches. We employed established metrics to evaluate the model's performance: Dice Similarity Coefficient (DSC) and Hausdorff Distance at the 95th percentile (HD95). These metrics quantify the overlap and spatial correspondence between auto-generated and ground-truth tumour contours. Utilizing patient 62 samples (9920 axial CT images) for training and 28 samples for independent testing, our model achieved a promising average DSC of 76.22% and HD95 of 8.74 mm. These results indicate that the model can predict highly concordant tumour volumes and generate sharp delineation boundaries, demonstrating remarkable agreement with ground truth data. Our initial finding shows the emerging potential of spectral-enhanced transformers for automated HNC tumour segmentation. This AI-powered approach holds significant promise for revolutionizing radiotherapy planning workflows. Streamlining time-intensive manual segmentation can optimize resource allocation for radiation oncologists, allowing them to focus on patient care and treatment planning & Potentially minimize inter-observer variability, leading to more consistent and potentially improved patient outcomes. Rajesh Pasricha (he/him/his), n/a (Presenting Author) - AIIMS Bhopal; Tanmay Basu (he/him/his), Assistant Professor (Co-Author) - IISERB; Vinod Kurmi (he/him/his), Assistant Professor (Co-Author) - IISERB; Avaneesh Mishra (he/him/his), RSO cum medical physicist (Co-Author) - AIIMS Bhopal; Manish Gupta (he/him/his), Professor (Co-Author) - AIIMS Bhopal; Saikat Das (he/him/his), Additional professor (Co-Author) - AIIMS Bhopal; Vipin Kharade (he/him/his), Assistant Professor (Co-Author) - AIIMS Bhopal; Aravind Padamanabhan (he/him/his), Senior Resident (Co-Author) - AIIMS Bhopal; B. Srinivas Reddy (he/him/his), PG Resident (Co-Author) - AIIMS Bhopal; Rajesh Malik (he/him/his), professor (Co-Author) - AIIMS Bhopal