import spaces import gradio as gr import torch from model import MedSegNet from inference import preprocess_volume, run_inference, make_slice_figure, make_combined_figure from report import build_pdf_report device = "cuda" if torch.cuda.is_available() else "cpu" model = MedSegNet( in_channels=4, sr_ratios=[4, 2, 1, 1], embed_dims=[32, 64, 160, 256], patch_kernel_size=[7, 3, 3, 3], patch_stride=[4, 2, 2, 2], patch_padding=[3, 1, 1, 1], mlp_ratios=[4, 4, 4, 4], num_heads=[1, 2, 5, 8], depths=[2, 2, 2, 2], decoder_head_embedding_dim=256, num_classes=3, decoder_dropout=0.0, ) state_dict = torch.load("best_model.pt", map_location="cpu") model.load_state_dict(state_dict) model.eval() @spaces.GPU def run(flair_file, t1_file, t1ce_file, t2_file): if not all([flair_file, t1_file, t1ce_file, t2_file]): raise gr.Error("Please upload all 4 modalities: FLAIR, T1, T1ce, and T2.") model.to("cuda") volume, affine = preprocess_volume(t1_file, t1ce_file, t2_file, flair_file) pred = run_inference(model, volume, "cuda") tc_voxels = int(pred[0].sum()) wt_voxels = int(pred[1].sum()) et_voxels = int(pred[2].sum()) stats_md = ( f"### Predicted Region Volumes (voxels)\n" f"- **Whole Tumor (WT):** {wt_voxels:,}\n" f"- **Tumor Core (TC):** {tc_voxels:,}\n" f"- **Enhancing Tumor (ET):** {et_voxels:,}\n" ) fig_path = make_slice_figure(volume, pred) combined_path = make_combined_figure(volume, pred) pdf_path = build_pdf_report(volume, pred, fig_path, combined_path) return combined_path, fig_path, stats_md, pdf_path def load_example(): return ( "examples/BRATS_001_0000.nii.gz", # FLAIR "examples/BRATS_001_0001.nii.gz", # T1 "examples/BRATS_001_0002.nii.gz", # T1ce "examples/BRATS_001_0003.nii.gz", # T2 ) theme = gr.themes.Soft( primary_hue="blue", secondary_hue="cyan", ).set( body_background_fill="*neutral_50", ) with gr.Blocks(title="MedSegNet — 3D Brain Tumor Segmentation", theme=theme) as demo: gr.Markdown("# MedSegNet") with gr.Row(): with gr.Column(scale=1): gr.Markdown("#### 1. Upload MRI Volumes") gr.Markdown("*(.nii or .nii.gz format, all 4 modalities required)*") flair = gr.File(label="FLAIR") t1 = gr.File(label="T1") t1ce = gr.File(label="T1ce") t2 = gr.File(label="T2") with gr.Row(): example_btn = gr.Button("📁 Load Example Case", variant="secondary") submit_btn = gr.Button("▶ Run Segmentation", variant="primary") pdf_output = gr.DownloadButton("📄 Download PDF Report", variant="secondary") with gr.Column(scale=1): gr.Markdown("#### 2. Segmentation Result") combined_image = gr.Image(label="Combined Segmentation Overlay") stats_output = gr.Markdown() gr.Markdown("#### Region Breakdown") output_image = gr.Image(label="FLAIR / WT / TC / ET") example_btn.click(fn=load_example, outputs=[flair, t1, t1ce, t2]) submit_btn.click( fn=run, inputs=[flair, t1, t1ce, t2], outputs=[combined_image, output_image, stats_output, pdf_output], ) demo.launch()