A Deep Learning-Based Computer Aided Detection (CAD) System for Difficult-to-Detect Brain Metastases (BM)

• Published in: International journal of radiation oncology, biology, physics Vol. 114; no. 3; pp. S111 - S112
• Main Authors: Fairchild, A., Salama, J.K., Wiggins, W., Ackerson, B., Fecci, P., Kirkpatrick, J.P., Floyd, S.R., Godfrey, D.J.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.11.2022
• ISSN: 0360-3016
• DOI: 10.1016/j.ijrobp.2022.07.546

After an initial course of SRS, new BMs are often identified in subsequent imaging, prompting additional SRS treatment.  Upon review, many are already subtly evident (enhancing) on MRIs from the first SRS. We refer to these lesions as "retrospectively identified metastases" (RIMs).  We sought to develop a CAD system that specifically excels at identifying small, subtle BMs that frequently go undetected by humans, by training a convolutional neural network (CNN) on a unique MRI dataset including many RIMs. All patients with BMs treated with SRS alone at our institution from 2016 to 2018 were eligible, except those with prior brain-directed therapy or small cell histology.  For patients who received two courses of SRS within 6 months, 3D axial post-contrast spoiled gradient recalled echo MRIs from the initial course were evaluated for RIMs.  RIMs were categorized based on whether they met (+DC) or did not meet (-DC) diagnostic imaging-based criteria to classify them as metastases.  Prospectively identified and treated metastases (PIMs) from these patients were also included, as were PIMs from patients who received a single SRS course and exhibited no "new" metastases for at least one year post SRS. Patients were randomized into 5 groups:  4 training/validation, 1 test.   A modified open-source CNN, DeepMedic (DM), was optimized and trained using data augmentation and 4-fold cross-validation.  Our final CAD solution employs ensembling, discrimination via a 3D conditional random field algorithm, and thresholding strategies.  CAD performance was evaluated on the test set. 135 patients with 563 metastases were included: 72 RIMs (45 +DC, 27 -DC) and 491 PIMs, with median diameters of 2.7mm (0.9, 11.0) and 6.7mm (0.9, 37.9), respectively.   CAD results are summarized in Table 1.  High sensitivity was achieved for all prospectively-treatable BMs (PIMs and +DC RIMs): 93% overall, 79% < 3mm.  80% of all +DC RIMs were identified. A median of only 2 false positives were generated per patient, and the mean Dice similarity coefficient was 0.79. A novel BM CAD system has been developed and trained using a unique MRI dataset including many RIMs.  The algorithm demonstrates excellent sensitivity, even for RIMs, along with high specificity.  These results outperform published works, especially for small BMs < 3mm. Study limitations include the modest number of patients from a single institution, retrospective nature, and selective inclusion criteria.  Prospective, multi-institutional validation of the approach is warranted.