Quality Assurance for AI-Based Applications in Radiation Therapy.

Recent advancements in artificial intelligence (AI) in the domain of radiation therapy (RT) and their integration into modern software-based systems raise new challenges to the profession of medical physics experts. These AI algorithms are typically data-driven, may be continuously evolving, and their behavior has a degree of (acceptable) uncertainty due to inherent noise in training data and the substantial number of parameters that are used in the algorithms. These characteristics request adaptive, and new comprehensive quality assurance (QA) approaches to guarantee the individual patient treatment quality during AI algorithm development and subsequent deployment in a clinical RT environment. However, the QA for AI-based systems is an emerging area, which has not been intensively explored and requires interactive collaborations between medical doctors, medical physics experts, and commercial/research AI institutions. This article summarizes the current QA methodologies for AI modules of every subdomain in RT with further focus on persistent shortcomings and upcoming key challenges and perspectives.

Classification of various sources of error in range assessment using proton radiography and neural networks in head and neck cancer patients.

This study evaluates the suitability of convolutional neural networks (CNNs) to automatically process proton radiography (PR)-based images. CNNs are used to classify PR images impaired by several sources of error affecting the proton range, more precisely setup and calibration curve errors. PR simulations were performed in 40 head and neck cancer patients, at three different anatomical locations (fields A, B and C, centered for head and neck, neck and base of skull coverage). Field sizes were 26 × 26cm2for field A and 4.5 × 4.5cm2for fields B and C. Range shift maps were obtained by comparing an unperturbed reference PR against a PR where one or more sources of error affected the proton range. CT calibration curve errors in soft, bone and fat tissues and setup errors in the anterior-posterior and inferior-superior directions were simulated individually and in combination. A CNN was trained for each type of PR field, leading to three CNNs trained with a mixture of range shift maps arising from one or more sources of range error. To test the full/partial/wrong agreement between predicted and actual sources of range error in the range shift maps, exact, partial and wrong match percentages were computed for an independent test dataset containing range shift maps arising from isolated or combined errors, retrospectively. The CNN corresponding to field A showed superior capability to detect isolated and combined errors, with exact matches of 92% and 71% respectively. Field B showed exact matches of 80% and 54%, and field C resulted in exact matches of 77% and 41%. The suitability of CNNs to classify PR-based images containing different sources of error affecting the proton range was demonstrated. This procedure enables the detection of setup and calibration curve errors when they appear individually or in combination, providing valuable information for the interpretation of PR images.