Reproducing RECIST lesion selection via machine learning: Insights into intra and inter-radiologist variation.

Background: The Response Evaluation Criteria in Solid Tumors (RECIST) aims to provide a standardized approach to assess treatment response in solid tumors. However, discrepancies in the selection of measurable and target lesions among radiologists using these criteria pose a significant limitation to their reproducibility and accuracy. This study aimed to understand the factors contributing to this variability. Methods: Machine learning models were used to replicate, in parallel, the selection process of measurable and target lesions by two radiologists in a cohort of 40 patients from an internal pan-cancer dataset. The models were trained on lesion characteristics such as size, shape, texture, rank, and proximity to other lesions. Ablation experiments were conducted to evaluate the impact of lesion diameter, volume, and rank on the selection process. Results: The models successfully reproduced the selection of measurable lesions, relying primarily on size-related features. Similarly, the models reproduced target lesion selection, relying mostly on lesion rank. Beyond these features, the importance placed by different radiologists on different visual characteristics can vary, specifically when choosing target lesions. Worth noting that substantial variability was still observed between radiologists in both measurable and target lesion selection. Conclusions: Despite the successful replication of lesion selection, our results still revealed significant inter-radiologist disagreement. This underscores the necessity for more precise guidelines to standardize lesion selection processes and minimize reliance on individual interpretation and experience as a means to bridge existing ambiguities.

How Does Target Lesion Selection Affect RECIST? A Computer Simulation Study.

OBJECTIVES: Response Evaluation Criteria in Solid Tumors (RECIST) is grounded on the assumption that target lesion selection is objective and representative of the change in total tumor burden (TTB) during therapy. A computer simulation model was designed to challenge this assumption, focusing on a particular aspect of subjectivity: target lesion selection. MATERIALS AND METHODS: Disagreement among readers and the disagreement between individual reader measurements and TTB were analyzed as a function of the total number of lesions, affected organs, and lesion growth. RESULTS: Disagreement rises when the number of lesions increases, when lesions are concentrated on a few organs, and when lesion growth borders the thresholds of progressive disease and partial response. There is an intrinsic methodological error in the estimation of TTB via RECIST 1.1, which depends on the number of lesions and their distributions. For example, for a fixed number of lesions at 5 and 15, distributed over a maximum of 4 organs, the error rates are observed to be 7.8% and 17.3%, respectively. CONCLUSIONS: Our results demonstrate that RECIST can deliver an accurate estimate of TTB in localized disease, but fails in cases of distal metastases and multiple organ involvement. This is worsened by the "selection of the largest lesions," which introduces a bias that makes it hardly possible to perform an accurate estimate of the TTB. Including more (if not all) lesions in the quantitative analysis of tumor burden is desirable.

Development of a Prognostic AI-Monitor for Metastatic Urothelial Cancer Patients Receiving Immunotherapy.

Background: Immune checkpoint inhibitor efficacy in advanced cancer patients remains difficult to predict. Imaging is the only technique available that can non-invasively provide whole body information of a patient's response to treatment. We hypothesize that quantitative whole-body prognostic information can be extracted by leveraging artificial intelligence (AI) for treatment monitoring, superior and complementary to the current response evaluation methods. Methods: To test this, a cohort of 74 stage-IV urothelial cancer patients (37 in the discovery set, 37 in the independent test, 1087 CTs), who received anti-PD1 or anti-PDL1 were retrospectively collected. We designed an AI system [named prognostic AI-monitor (PAM)] able to identify morphological changes in chest and abdominal CT scans acquired during follow-up, and link them to survival. Results: Our findings showed significant performance of PAM in the independent test set to predict 1-year overall survival from the date of image acquisition, with an average area under the curve (AUC) of 0.73 (p < 0.001) for abdominal imaging, and 0.67 AUC (p < 0.001) for chest imaging. Subanalysis revealed higher accuracy of abdominal imaging around and in the first 6 months of treatment, reaching an AUC of 0.82 (p < 0.001). Similar accuracy was found by chest imaging, 5-11 months after start of treatment. Univariate comparison with current monitoring methods (laboratory results and radiological assessments) revealed higher or similar prognostic performance. In multivariate analysis, PAM remained significant against all other methods (p < 0.001), suggesting its complementary value in current clinical settings. Conclusions: Our study demonstrates that a comprehensive AI-based method such as PAM, can provide prognostic information in advanced urothelial cancer patients receiving immunotherapy, leveraging morphological changes not only in tumor lesions, but also tumor spread, and side-effects. Further investigations should focus beyond anatomical imaging. Prospective studies are warranted to test and validate our findings.

Prognostic Value of Deep Learning-Mediated Treatment Monitoring in Lung Cancer Patients Receiving Immunotherapy.

BACKGROUND: Checkpoint inhibitors provided sustained clinical benefit to metastatic lung cancer patients. Nonetheless, prognostic markers in metastatic settings are still under research. Imaging offers distinctive advantages, providing whole-body information non-invasively, while routinely available in most clinics. We hypothesized that more prognostic information can be extracted by employing artificial intelligence (AI) for treatment monitoring, superior to 2D tumor growth criteria. METHODS: A cohort of 152 stage-IV non-small-cell lung cancer patients (NSCLC) (73 discovery, 79 test, 903CTs), who received nivolumab were retrospectively collected. We trained a neural network to identify morphological changes on chest CT acquired during patients' follow-ups. A classifier was employed to link imaging features learned by the network with overall survival. RESULTS: Our results showed significant performance in the independent test set to predict 1-year overall survival from the date of image acquisition, with an average area under the curve (AUC) of 0.69 (p < 0.01), up to AUC 0.75 (p < 0.01) in the first 3 to 5 months of treatment, and 0.67 AUC (p = 0.01) for durable clinical benefit (6 months progression-free survival). We found the AI-derived survival score to be independent of clinical, radiological, PDL1, and histopathological factors. Visual analysis of AI-generated prognostic heatmaps revealed relative prognostic importance of morphological nodal changes in the mediastinum, supraclavicular, and hilar regions, lung and bone metastases, as well as pleural effusions, atelectasis, and consolidations. CONCLUSIONS: Our results demonstrate that deep learning can quantify tumor- and non-tumor-related morphological changes important for prognostication on serial imaging. Further investigation should focus on the implementation of this technique beyond thoracic imaging.