Quantitative results of our method with different modifications. This table illustrates the performance of our model with the removal of each component, assessed using both pixel- and image-level metrics. <i>Ours w/o Agg.</i> refers to our model without the part the slices’ feature aggregating component, essentially transforming it into a 2D version that processes brain MRI data slice by slice. <i>Ours w/o EDC</i> indicates that the feature extractor was not fine-tuned, with the extracted features directly applied to anomaly detection tasks. <i>SimpleNet w/ Agg.</i> represents a modification where our CNF is replaced with a binary classification model, effectively making it a 3D version of SimpleNet that utilizes our aggregation approach. Finally, <i>Ours</i> denotes our proposed method with Triplet Loss.

Quantitative results of our method with different modifications. This table illustrates the performance of our model with the removal of each component, assessed using both pixel- and image-level metrics. Ours w/o Agg. refers to our model without the part the slices’ feature aggregating component, essentially transforming it into a 2D version that processes brain MRI data slice by slice. Ours w/o EDC indicates that the feature extractor was not fine-tuned, with the extracted features directly applied to anomaly detection tasks. SimpleNet w/ Agg. represents a modification where our CNF is replaced with a binary classification model, effectively making it a 3D version of SimpleNet that utilizes our aggregation approach. Finally, Ours denotes our proposed method with Triplet Loss.