Overcoming challenges of translating deep-learning models for glioblastoma: the ZGBM consortium

Abstract

Objective: To report imaging protocol and scheduling variance in routine care of glioblastoma patients in order to demonstrate challenges of integrating deep learning models in glioblastoma care pathways. Additionally, to understand the most common imaging studies and image contrasts to inform the development of potentially robust deep learning models. Methods: MR imaging data were analysed from a random sample of 5 patients from the prospective cohort across five participating sites of the ZGBM consortium. Reported clinical and treatment data alongside DICOM header information were analysed to understand treatment pathway imaging schedules. Results: All sites perform all structural imaging at every stage in the pathway except for the presurgical study, where in some sites only contrast-enhanced T1-weighted imaging is performed. Diffusion MRI is the most common non-structural imaging type, performed at every site. Conclusion: The imaging protocol and scheduling varies across the UK, making it challenging to develop machine learning models that could perform robustly at other centres. Structural imaging is performed most consistently across all centres. Advances in knowledge: Successful translation of deep learning models will likely be based on structural post-treatment imaging unless there is significant effort made to standardise non-structural or peri-operative imaging protocols and schedules.

Glioblastoma and radiotherapy: A multicenter AI study for Survival Predictions from MRI (GRASP study)

“… 6 While not the focus of our study which incorporates all patients consecutively (including complete response, partial response, stable disease, progression, and pseudoprogression), interpreting post-radiotherapy structural MRIs in clinical settings is typically challenging due to difficulty in distinguishing recurrent disease from treatment-related effects—particularly for pseudoprogression. 2 , 3 , 6 , 11 , 17 , 36–38 However, labelling progression—and pseudoprogression—requires the availability of repeated T1c imaging obtained in a timely manner per patient, accompanied by accurate measurements of bidirectional diameters of contrast-enhancing tumors. 37 , 38 Prior research has reported that there can be substantial inter-rater variability in these measurements, however, which can confound evaluations of treatment response.…”

“…These sequences were consistently acquired at all centers; conversely, more advanced MRIs are less commonly available. 17 , 18 Incorporating other anatomical sequences desirable for brain tumor imaging, such as FLAIR sequences, was not also pursued as it would have reduced the patient cohort in this UK-based study where FLAIR imaging was not always performed. A downstream constraint of building a model without the most common MRI sequences is that it reduces the potential for clinical translation.…”

“…Patient cohort .—This study included consecutive retrospective and prospective data from 11 ZGBM (zeugmatography for glioblastoma) consortium centers, 17 with diagnoses between March 2014 and February 2022 (a CONSORT diagram displaying the flow of patients included in analyses is presented in Supplementary Appendix C ). The study was pragmatic; imaging regimens were not standardized and were expected to vary over centers and time.…”

Application value of T2 fluid-attenuated inversion recovery sequence based on deep learning in static lacunar infarction

“…It has been previously reported that deep learning AI sequences can reduce scanning time by 45% − 60% while keeping spatial resolution constant (29–31). This improves patient experiences and scanner efficiency without sacrificing diagnostic quality, which is consistent with the findings of the present study.…”