S2302
Physics - Machine learning and AI algorithms
ESTRO 2026
EJ, Shen Y, Wallis P, Allen-Zhu Z, Li Y, Wang S, et al. LoRA: Low-Rank Adaptation of Large Language Models. arXiv; 2021. Available from: http://arxiv.org/abs/2106.09685 Keywords: Foundational models, resource-limited settings Digital Poster 3068 Comparison of AI based techniques for toxicity prediction Antony Carver 1 , Andrew Hartley 2 , Janet Dunn 3 , Hisham Mehanna 4 1 Deparment of Medical Physics, University Hospital Birmingham, Birmingham, United Kingdom. 2 Department of Oncology, University Hospital Birmingham, Birmingham, United Kingdom. 3 Warwick Clinical Trials Unit, University of Warwick, Coventry, United Kingdom. 4 Institute for Head and Neck Studies and Education, University of Birmingham, Birmingham, United Kingdom Purpose/Objective: Machine learning based techniques such as CNNs (Convolutional Neural Networks) have shown great potential for learning features in images predictive of outcomes. Outcome data is often imperfect, with missing and dubious values. This can lead to overfitting and poor generalisation. Moreover, interpreting models can be challenging. An alternative approach is a hierarchical model, where the output of the machine learning component is the parameters of a simple model. This guarantees the result takes an assumed form and makes interpretation much easier as the weights of the model are tied to a clear feature of the patient recovery. This study compares how different approaches to using machine learning to predict the time course of toxicity perform in the presence of noisy and in-consistent real-word data. Material/Methods: Overall MDADI (MD Anderson Dysphagia Inventory) [1] scores from 187 patients treated as part of the DeEscalate trial [2] were recorded at baseline, post- treatment, 3-months, 6-months and 12-months post- treatment. A recovery curve model was defined to predict the overall MDADI score, S, for up to a year after treatment. The three-parameter model,S(t) = AS0+B(1 – exp(-Ct))(1-AS0)The parameter A describes an initial drop in MDADI following treatment, followed by a final recovery controlled by parameter, B with rate-of-recovery C. The MDADI at time, t, was predicted using the dose grid, planning images and baseline MDADI score, S0.Four probabilistic deep neural networks were compared to a simple fit of the data to the recovery curve (Figure 1). The first two were based on a CNN-MLP (Multi-Layer Perceptron)
the RankMe score [3], and with online linear probing.To compare classification performance, we used a simple downstream task of identifying whether a CBCT image contained a nasogastric feeding tube in an expert-annotated subset of the pre-training dataset. We compared the pre-trained model fine- tuned using low-rank adapters (LoRA) [4] and a trained-from-scratch model using supervised learning.All experiments were run on RTX3090 cards. Results: In Figure 1a, the L1 validation loss was lower for ViT- small at the expense of a higher computing resources (peta FLOPS days). The RankMe score (Figure 1b) and the online linear probing’s AUROC (Figure 1c) increased with training iterations. In Table 1, the final test set performance is shown for the binary classification downstream task, where LoRA reached the highest value for all reported metrics.
Conclusion: It is possible to train a foundational model in an RT department with a resource-limited compute, by effective training tracking and informed hyperparameter selection based on training curves. A better classification performance was obtained with LoRA on a down-stream task compared to training from scratch on a large dataset of 48,500 patients, supporting the relevance of pre-trained image models in future applications. References: 1.Haghighi F, Hosseinzadeh Taher MR, Gotway MB, Liang J. Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial? Med Image Anal. 2024 May;94:103086. 2.Xie Z, Zhang Z, Cao Y, Lin Y, Bao J, Yao Z, et al. SimMIM: A Simple Framework for Masked Image Modeling. arXiv; 2022. Available from: http://arxiv.org/abs/2111.098863.Garrido Q, Balestriero R, Najman L, Lecun Y. RankMe: Assessing the downstream perfor-mance of pretrained self- supervised representations by their rank. arXiv; 2023. Available from: http://arxiv.org/abs/2210.028854.Hu
Made with FlippingBook - Share PDF online