CT Head Hemorrhage Detection Research model for slice and series level intracranial hemorrhage detection and slice level localization on noncontrast head CT. The complete system combines: 1. a 9 channel MaxViT Tiny slice encoder/classifier; 2. a two layer bidirectional GRU for contextual slice and series predictions; 3. five DeepLabV3+ decoders sharing that exact encoder state for radiologist facing localization. This is research software, not a medical device . It must not be used to diagnose, exclude, triage, or manage intracranial hemorrhage without independent clinical validation and appropriate regulatory review. Model Details Component Specification Slice encoder maxvit tiny tf 512.in1k , ImageNet initialization Slice input Three adjacent axial slices centered on the predicted slice Channels Brain, subdural, and bone windows for each slice, flattened to 9 channels Slice outputs Epidural, intraparenchymal, intraventricular, subarachnoid, subdural, and any hemorrhage Sequence head Two layer bidirectional GRU, 256 hidden units/direction, class specific attention Sequence outputs Context refined slice probabilities and attention pooled series probabilities Localization Five fold…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy