Two-Step Semi-Automated Classification of Choroidal Metastases on MRI: Orbit Localization via Bounding Boxes Followed by Binary Classification via Evolutionary Strategies

Abstract

Background and purpose: The choroid of the eye is a rare site for metastatic tumor spread, and as small lesions on the periphery of brain MRI studies, these choroidal metastases are often missed. To improve their detection, we aimed to use artificial intelligence to distinguish between brain MRI scans containing normal orbits and choroidal metastases.

Materials and methods: We present a novel hierarchical deep learning framework for sequential cropping and classification on brain MRI images to detect choroidal metastases. The key innovation of this approach lies in training an orbit localization network based on a YOLOv5 architecture to focus on the orbits, isolating the structures of interest and eliminating irrelevant background information. The initial sub-task of localization ensures that the input to the subsequent classification network is restricted to the precise anatomical region where choroidal metastases are likely to occur. In Step 1, we trained a localization network on 386 T2-weighted brain MRI axial slices from 97 patients. Using the localized orbit images from Step 1, in Step 2 we trained a binary classifier network with 33 normal and 33 choroidal metastasis-containing brain MRIs. To address the challenges posed by the small dataset, we employed a data-efficient evolutionary strategies approach, which has been shown to avoid both overfitting and underfitting in small training sets.

Results: Our orbit localization model identified globes with 100% accuracy and a mean Average Precision of Intersection over Union thresholds of 0.5 to 0.95 (mAP(0.5:0.95)) of 0.47 on held-out testing data. Similarly, the model generalized well to our Step 2 dataset which included orbits demonstrating pathologies, achieving 100% accuracy and mAP(0.5:0.95) of 0.44. mAP(0.5:0.95) appeared low because the model could not distinguish left and right orbits. Using the cropped orbits as inputs, our evolutionary strategies-trained convolutional neural network achieved a testing set area under the curve (AUC) of 0.93 (95% CI [0.83, 1.03]), with 100% sensitivity and 87% specificity at the optimal Youden's index.

Conclusions: The semi-automated pipeline from brain MRI slices to choroidal metastasis classification demonstrates the utility of a sequential localization and classification approach, and clinical relevance for identifying small, "corner-of-the-image", easily overlooked lesions.

Abbreviations: AI = artificial intelligence; AUC = area under the curve; CNN = convolutional neural network; DNE = deep neuroevolution; IoU = intersection over union; mAP = mean average precision; ROC = receiver operating characteristic.