An Open-Access Deep Learning Tool for Early Detection of Chiari Type I Malformation via MRI Classification

Chiari Type I Malformation (CM-1) is a congenital neurological condition in which the cerebellar tonsils descend below the foramen magnum, compressing the brainstem and disrupting cerebrospinal fluid dynamics. In Colombia and much of Latin America, the scarcity of specialized neuroradiology expertise outside major urban centers leads to diagnostic delays that can span years, limiting access to timely surgical intervention. We present a complete, reproducible pipeline---from open-access dataset construction to public web deployment---for automated CM-1 detection in mid-sagittal T1-weighted brain MRI. Our system applies two-phase transfer learning on DenseNet121, EfficientNetB0, and ResNet50V2, trained on a curated dataset of 156 real clinical images (126 CM-1, 30 normal) obtained under a CC BY-NC-SA 4.0 license and preprocessed with CLAHE contrast enhancement. Evaluated on an independent external test set of 131 images, DenseNet121 achieved AUC-ROC 0.943, sensitivity 0.852, specificity 0.933, and F1-score 0.910. Gradient-weighted Class Activation Mapping (Grad-CAM) confirms that the model attends to the posterior cranial fossa and foramen magnum---the anatomical structures that define CM-1 radiologically. The tool is publicly deployed, requires no registration, runs inference in under one second, and implements privacy-by-design by never writing images to disk. To the best of our knowledge, this is the first freely accessible deep-learning diagnostic aid for CM-1 developed in and for a Latin American context.