Abstract
SparseProto combines global convolutional features with a top-k prototype-assignment histogram in a biomedical image classifier with 41,187 to 42,759 parameters. We evaluate predictive performance, parameter count, counted operations, and measured latency on BloodMNIST, PathMNIST, and DermaMNIST. Explainability is not evaluated. The original three-seed comparison illustrates that parameter and operation reductions do not translate proportionally into latency reductions. The prototype assignment is sparse only after a dense similarity computation and therefore does not establish computational sparsity. In a separate five-seed rerun, the full model differed from a convolutional neural network (CNN)-only variant by +0.38, −2.21, and +0.44 percentage points on the three datasets, while on PathMNIST its graphics processing unit (GPU) latency (single-precision floating-point arithmetic, FP32; batch size 1) was 2.45 times higher. The available evidence does not establish a consistent predictive advantage or a latency benefit of adding the prototype branch.