A Multi-Centric Evaluation of AI-Driven OARs Low Field MRgRT Pelvic /Abdomen Contouring

Abstract

<h3>Purpose/Objective(s)</h3> Organs at risk annotation is a strong bottleneck of Magnetic resonance imaging guided radiotherapy (MRgRT) in the context of adaptive treatment. It is a time-consuming task that reduces patient throughput (20% of the fraction duration dedicated to contouring) while suffering standardization and reproducibility across physicians, hampering the accuracy of high precision MRgRT and diminishing its adoption potential. AI-contouring becomes a game changer in radiation oncology since it is able within seconds to provide a full OAR delineation that could be close to clinical acceptance with little modifications. The aim of this study is to evaluate the performance of AI-contouring within a multi-centric cohort for patients with pelvic / abdominal tumors treated with low field (0.35T) MRgRT. <h3>Materials/Methods</h3> In the context of this study, a CE/FDA cleared anatomically preserving ensemble deep-learning architecture contouring solution was considered. The adopted solution was trained using more than 350 0.35T MR fully annotated pelvic cases according to the ESTRO guidelines and 270 annotated abdomen fractions samples. A retrospective cohort involving 42 test cases coming from seven different institutions (US: 1, EU: 5, AS: 1) was considered. The clinical delineations used for treatment planning from expert physicians/medical physicists were associated with these cases. <h3>Results</h3> It appears that treatment practices can be very different between institutions since the use of OAR constraints were far from being uniformly adopted. Bladder & liver dosimetry constraints were the most frequently used (100% & 90%) while abdominal aorta and seminal vesicle were the least adopted (24% & 15%). The average DSC between Clinical experts and AI annotations was 78% across all structures. Bladder and left/right kidney were the structures for which the highest DSC were observed (93%, 91% & 90%), while penile bulb and duodenum were the ones with the lowest agreement (54% & 59%). AI solutions seem to have important discrepancies with clinical contours in organs on which either the volume is small or there are practice-related uncertainties with respect to the definition of beginning and the end of the structure. For quantitative evaluation, dice similarity coefficients (DSC) and 95% Hausdorff distances (HD95) were calculated. <h3>Conclusion</h3> This retrospective multi-centric study demonstrates that AI-driven contours could be a reliable alternative to clinical contours offering performance that appears to be close to the human expert for many of the structures while increasing throughput and offering automatization & standardization.