[Paper Review] Intentional deep overfit learning (IDOL): A novel deep learning strategy for adaptive radiation therapy

Jaehee Chun|arXiv (Cornell University)|Jan 1, 2022

Advanced Radiotherapy TechniquesPhysics and Astronomy20 references32 citations

TL;DR

This paper introduces Intentional Deep Overfit Learning (IDOL), a novel deep learning framework that intentionally overfits a model to patient-specific prior data to enhance performance in adaptive radiation therapy (ART). By first training a generalized model on diverse data and then fine-tuning it on a single patient’s augmented prior information, IDOL achieves significant performance gains—improving Dice scores by 0.088, reducing MRI super-resolution MAE by 40%, and cutting synthetic CT MAE by 65%—demonstrating its potential for personalized, data-efficient ART applications.

ABSTRACT

Purpose: Applications of deep learning (DL) are essential to realizing an effective adaptive radiotherapy (ART) workflow. Despite the promise demonstrated by DL approaches in several critical ART tasks, there remain unsolved challenges to achieve satisfactory generalizability of a trained model in a clinical setting. Foremost among these is the difficulty of collecting a task-specific training dataset with high-quality, consistent annotations for supervised learning applications. In this study, we propose a tailored DL framework for patient-specific performance that leverages the behavior of a model intentionally overfitted to a patient-specific training dataset augmented from the prior information available in an ART workflow-an approach we term Intentional Deep Overfit Learning (IDOL). Methods: Implementing the IDOL framework in any task in radiotherapy consists of two training stages: (1) training a generalized model with a diverse training dataset of N patients, just as in the conventional DL approach, and (2) intentionally overfitting this general model to a small training dataset-specific the patient of interest (N+1) generated through perturbations and augmentations of the available task- and patient-specific prior information to establish a personalized IDOL model. The IDOL framework itself is task-agnostic and is, thus, widely applicable to many components of the ART workflow, three of which we use as a proof of concept here: the autocontouring task on replanning CTs for traditional ART, the MRI super-resolution (SR) task for MRI-guided ART, and the synthetic CT (sCT) reconstruction task for MRI-only ART. Results: In the replanning CT autocontouring task, the accuracy measured by the Dice similarity coefficient improves from 0.847 with the general model to 0.935 by adopting the IDOL model. In the case of MRI SR, the mean absolute error (MAE) is improved by 40% using the IDOL framework over the conventional model. Finally, in the sCT reconstruction task, the MAE is reduced from 68 to 22 HU by utilizing the IDOL framework. Conclusions: In this study, we propose a novel IDOL framework for ART and demonstrate its feasibility using three ART tasks. We expect the IDOL framework to be especially useful in creating personally tailored models in situations with limited availability of training data but existing prior information, which is usually true in the medical setting in general and is especially true in ART.

Motivation & Objective

• To address the challenge of limited, high-quality training data in adaptive radiation therapy (ART) by leveraging patient-specific prior information.
• To overcome the limitations of conventional deep learning, which prioritizes generalization over patient-specific performance.
• To propose a new paradigm in medical AI that intentionally uses overfitting as a tool for personalization in precision medicine.
• To validate the IDOL framework across three critical ART tasks: auto-contouring, MRI super-resolution, and synthetic CT generation.

Proposed method

• The IDOL framework consists of two training stages: first, training a generalized model on a diverse dataset of N patients.
• Second, intentionally overfitting the generalized model to a small, patient-specific dataset derived from perturbations and augmentations of the patient’s prior imaging and contour data.
• The framework is task-agnostic and can be applied to any image-to-image translation or segmentation task in radiotherapy.
• It leverages existing prior knowledge—such as planning CTs, contours, or diagnostic scans—by generating augmented training samples to improve model fidelity for a single patient.
• The method uses standard deep learning architectures (e.g., U-Net variants) with a loss function optimized for task-specific metrics like Dice or MAE.
• The approach avoids hyperparameter tuning per task, relying instead on the model’s ability to memorize and exploit patient-specific patterns through controlled overfitting.

Experimental results

Research questions

• RQ1 Can intentional overfitting to patient-specific prior data significantly improve deep learning model performance in adaptive radiation therapy tasks?
• RQ2 Does the IDOL framework outperform conventional generalization-focused deep learning models in low-data regimes common in clinical radiotherapy?
• RQ3 To what extent can prior imaging and contour data be augmented to create effective patient-specific training sets for deep learning?
• RQ4 Is the IDOL framework generalizable across diverse ART tasks such as auto-contouring, MRI super-resolution, and synthetic CT generation?
• RQ5 Can IDOL achieve high performance without task-specific hyperparameter tuning, making it clinically deployable?

Key findings

• The IDOL model improved the Dice similarity coefficient (DSC) from 0.847 ± 0.031 to 0.935 ± 0.026 in the auto-contouring task on re-planning CTs.
• In MRI super-resolution, the mean absolute error (MAE) was reduced from 4.18 ± 0.86 to 2.50 ± 0.54, a 40% improvement.
• For synthetic CT generation, the MAE decreased from 68 ± 16 HU to 22 ± 5 HU, representing a 67% reduction.
• The IDOL framework achieved these results without task-specific hyperparameter tuning, demonstrating broad applicability.
• Training time for the personalized model was minimal—averaging 20 seconds for auto-contouring, 93 seconds for MRI SR, and 3 hours for sCT generation.
• The method showed consistent performance gains across all three tasks, validating its feasibility as a patient-specific deep learning strategy in ART.