S1611
Physics - Autosegmentation
ESTRO 2026
image transformer (SMIT). In MICCAI.[2] Hendrycks, D., Mazeika, M. and Dietterich, T. Deep anomaly detection with outlier exposure. Proceedings of ICLR, 2019.[3] Hendrycks, D. and Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks. Proceedings of ICLR, 2017.[4] Hendrycks D., Basart S., Mazeika M., Zou A., Mostajabi
x 4 mm spacing) of patients with locally advanced rectal cancer in 3-fold cross-validation and evaluated on a held-out set of 240 cases. nnU-Net was trained using standard protocol, whereas the transformer models were subject to MR standardization with 95th percentile intensity clipping [3]. Accuracy was computed using Dice similarity coefficient (DSC), 95th percentile Hausdroff distance (HD95), and contour acceptability (DSC > 0.7). Paired Wilcoxon signed-rank tests measured statistical differences between RectSegmentor and the other models. Results:
M., Steinhardt J., and Song D. Scaling out-of- distribution detection for real-world settings.
Proceedings of ICML, 2022.[5] Liu W., Wang X., Owens J., and Li Y. Energy-based out-of-distribution detection. NeurIPS 2020. Keywords: out-of-distribution, segmentation, transformers
Digital Poster 4793
Anisotropic resolution training improves transformer- based rectal tumor segmentation in oblique 3D MRI scans Aneesh Rangnekar 1 , Aditya Apte 1 , Eve LoCastro 1 , Paul Romessor 2,3 , Jesse Joshua Smith 4 , Julio Garcia-Aguilar 4 , Joseph Deasy 1 , Harini Veeraraghavan 1 1 Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA. 2 Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, USA. 3 Department of Medicine, Memorial Sloan Kettering Cancer Center, New York, USA. 4 Department of Surgery, Memorial Sloan Kettering Cancer Center, New York, USA Purpose/Objective: Rectal cancer is the third most common cancer globally, with rising incidence particularly in younger populations. MRI plays a critical role in staging, treatment planning, and surveillance using watch-and- wait to early detect local tumor regrowth. Methods to automatically and accurately extract tumor volumes would benefit radiomic-based biomarker studies as well as longitudinal assessment of tumor treatment response. Hence, deep learning models consisting of pretrained transformer encoder with U-Net decoder and convolutional network nnU-Net were analyzed for segmentation of rectal cancers from T2-weighted MRI. Material/Methods: A hybrid transformer encoder U-Net decoder network and nnU-Net [1] were used. The transformer was pretrained with the SMIT framework [2] using 10,432 unlabeled 3D CT scans of patients with various diseases involving head to pelvis. Unlike SMIT that uses isotropic resolution (128 x 128 x 128 voxels), RectSegmentor handles smaller field-of-view and avoids lossy interpolation or suboptimal padding strategies in MRI by using anisotropic resolution of 128 x 128 x 64 voxels. All models were trained with 167 oblique T2W MRI scans (1.5T and 3T, average 0.8 x 0.8
Table 1 shows that all three models were similarly accurate in cross-validation (SMIT p = 0.43, nnU-Net p = 0.69) but showed difference in testing (SMIT p = 0.005, nnU-Net p = 0.014) with respect to RectSegmentor. RectSegmentor also produced the highest percentage of clinically acceptable segmentations on both datasets. Figure 1 shows representative examples that show the differences in the segmentation of the three models with the last row depicting cases where all three models failed due to the presence of small lesions and poor soft-tissue contrast. Conclusion:
Made with FlippingBook - Share PDF online