Li, Ian D.

Background/Objectives: Glioblastoma surveillance after surgery and chemoradiation remains challenging because MRI findings of tumor recurrence can overlap with pseudo-progression, treatment-related effects, and postoperative tissue changes. Methods: We developed a seven-channel postoperative MRI framework using a deep learning transfer method to support non-invasive glioblastoma treatment-effect assessment. The model used T1 contrast-enhanced, FLAIR, T1, RSI-Cell, ADC, T2, and cerebral blood flow volumes from 124 postoperative glioblastoma patients (164 MRI timepoints) as input. Images were processed using a 3D ResNet18 encoder pretrained on 588 postoperative glioma samples and fine-tuned using task-specific classification heads. The cohort comprised 124 patients contributing 164 postoperative MRI timepoints, all acquired at 3T on scanners from a single vendor. Performance was evaluated with nested five-fold cross-validation stratified and assigned at the patient level, so that all timepoints from a given patient fell in one-fold and the training epoch was selected on an inner split rather than on the fold being reported. Because some of the clinical labels were incomplete, the number of evaluable timepoints differed by task (recurrence versus pseudo-progression, 164; MGMT, 99; short-term survival, 139). The whole procedure was repeated under three independent random seeds and results are reported as the mean and standard deviation across seeds. Results: The strongest clinical endpoint was recurrence versus pseudo-progression, where nested cross-validation across three random seeds gave a pooled out-of-fold AUC of 0.935 (SD = 0.014), area under the precision-recall curve of 0.973, balanced accuracy of 0.880, sensitivity of 0.917, and specificity of 0.843. No other endpoint reached reliable discrimination. Radiation decision reached an AUC of 0.658 (SD = 0.039), while MGMT promoter methylation (AUC = 0.532, SD = 0.082) and short-term survival (AUC = 0.514, SD = 0.027) were indistinguishable from chance. Conclusions: In this single-center proof-of-concept study, postoperative MRI successfully distinguished tumor recurrence from pseudo-progression. However, the models did not reliably predict the other three outcomes related to molecular status, treatment planning, and prognosis. Overall, the model learned imaging features specifically associated with recurrence, rather than a more general representation of the tumor that can predict many different clinical outcomes. These findings support technical feasibility for a single endpoint rather than clinical readiness, for which external multi-center validation is required.

Accurate estimation of Alzheimer’s disease (AD) severity from structural magnetic resonance imaging (MRI) remains difficult, as disease-associated anatomical alterations are often subtle and publicly available datasets are typically too small to support robust deep learning model training. This feasibility study sought to determine how much Alzheimer’s disease spectrum-related information could be extracted from a small structural MRI cohort using a deliberately lightweight two-dimensional convolutional neural network (2D CNN), and whether transfer learning improves model performance. This study was intended as a methodological proof of concept rather than the development of a clinically deployable diagnostic tool. Structural scans and Clinical Dementia Rating (CDR) labels from the OASIS-1 dataset were filtered to 214 subjects: 124 cognitively normal (CN), 65 with mild cognitive impairment (MCI; CDR = 0.5), and 25 with AD-level impairment (CDR ≥ 1). A compact 2D CNN trained from scratch and a transfer learning model (frozen ImageNet MobileNetV2 features) were evaluated on four binary tasks (CN vs. AD, MCI vs. AD, CN vs. MCI, and CN vs. any impairment) under identical pre-processing and subject-level repeated 5-fold cross-validation (10 repeats), with the decision threshold tuned only on an inner split. Discrimination was summarized by ROC-AUC with 95% confidence intervals (CIs), permutation tests against chance, and per-task sensitivity and specificity. The from-scratch CNN recovered only a broad normal-versus-impaired signal (CN vs. any impairment AUC 0.59) and was at chance on adjacent-stage tasks (MCI vs. AD 0.41; CN vs. MCI 0.51). Transfer learning improved every task: CN vs. AD AUC 0.745 (95% CI 0.730–0.763), CN vs. any impairment 0.642, CN vs. MCI 0.601, and MCI vs. AD 0.599. On an independent OASIS-2 cohort, the transfer learning CN vs. AD model retained AUC 0.748. In this small-data regime, transfer learning recovers substantially more Alzheimer-spectrum signals than a from-scratch CNN, but performance remains modest because it is bounded by CDR-based, non-biomarker-confirmed labels, suggesting the model separates CDR-defined cognitive-status groups rather than detecting AD pathology.