2025 Poster Presentations
P343: DEEP LEARNING FOR AUTOMATIC SEGMENTATION OF PITUITARY ADENOMAS: A VIDEOMICS STUDY
Edoardo Agosti1; Alberto Paderno2; Andrea Pagnoni1; Vittorio Rampinelli1; Pier Paolo Panciani1; Alessandro Fiorindi1; Cesare Piazza1; Marco Maria Fontanella1; 1University of Brescia; 2Humanitas University of Rozzano
Introduction: Fully convolutional neural networks (FCNNs) applied to video analysis are of particular interest in the field of head and neck oncology. The application of video analysis to diagnostic endoscopy has been termed 'videomics'. Recently, videomics has also been applied to endonasal endoscopy, but to our knowledge, no study has validated the possible role, advantages, and limitations of videomics in the intraoperative recognition of pituitary adenomas (PAs). The aim of this study was to test methods based on FCNNs for the semantic segmentation of PAs.
Methods: A dataset was retrieved from the institutional registry of a tertiary academic hospital. The inclusion criteria were adult patients with PA, who underwent endoscopic endonasal surgery for the first time with narrow band imaging (NBI) endoscopic video recording. Manual segmentation of video frames was performed with Label Studio software (Figure 1). Three FCNNs (U-Net, U-Net 3, and ResNet) were studied for segmenting the PA images. The performance of the FCNNs was evaluated for each network tested and compared with the gold standard, represented by manual annotation performed by expert physicians.
Results: A total of 20 NBI endoscopic videos of PAs surgery has been analyzed. For each video, 35 frames were manually selected by two authors. The dataset consisted of 700 frames. The best results in terms of the Dice similarity coefficient (Dsc) were obtained by ResNet with 5 blocks (× 2) and 16 filters, with an average value of 0.6559. All tested FCNNs exhibited very high variance values, leading to very low minimum values for all evaluated parameters. The inference time of the processing networks ranged from 14 ms to 115 ms.
Conclusion: FCNNs, in particular ResNet, show promising potential in the analysis and segmentation of video-endoscopic images of PAs. All tested FCNNs architectures demonstrated satisfactory results in terms of diagnostic accuracy. The inference time of the processing networks was particularly short, thus showing the possibility of real-time application.
