Evaluating the Efficacy of Deep Learning Models for Distinguishing Spinal Hemangiomas from Common Metastatic Tumors in Across MRI Sequences

Purpose This study aimed to develop and evaluate a deep learning approach for distinguishing spinal hemangiomas from thoracolumbar spinal metastases (breast, lung, prostate, renal cell carcinoma, and multiple myeloma) across MRI sequences and assess sequence-dependent sensitivity. Method Radiological data from a university hospital (2016–2023) were collected from 352 adults (mean age 61.3 years): 109 hemangiomas and 243 metastases (histopathologically confirmed). Lesions were labeled and regions of interest annotated using VGG Image Annotator, cropped, converted to grayscale, and resized to 256×256 pixels across four sequences (T1, T2, STIR, contrast-enhanced). A unified CNN (three convolutional layers with 32/64/128 filters, max pooling, batch normalization, 128-neuron fully connected layer, dropout 0.5, softmax output) was trained with downsampling-based class balancing and a 60/40 stratified train–test split (Adam, categorical cross-entropy, 40 epochs, batch size 16), with hyperparameter optimization over learning rate, batch size, epochs, and dropout. Performance varied markedly by sequence and comparison. Results T1 and T2 yielded the most robust discrimination; the best result occurred on T1 for hemangioma versus multiple myeloma (accuracy 0.89, AUC 0.93). Across all T2 datasets, performance was balanced and high (accuracy 0.84, sensitivity 0.93, specificity 0.75, AUC 0.87). In contrast, STIR and contrast-enhanced sequences showed reduced and inconsistent performance (overall STIR accuracy 0.57, AUC 0.54; contrast-enhanced accuracy 0.68, AUC 0.72), including occasional near-random classification and pronounced sensitivity–specificity imbalances. Conclusion Deep learning-based differentiation of hemangiomas from spinal metastases is feasible using conventional MRI, with T1/T2 sequences providing reliable features, while STIR and post-contrast imaging may need larger datasets and advanced architectures to improve robustness.