DeepSAM

Code and experiments accompanying the article "DeepSAM - Self-Awareness as a Way to Mitigate Manifestation Shifts in Medical Images" .

DeepSAM is an unsupervised anomaly detection framework for medical images. It uses a convolutional autoencoder trained only on normal samples and detects anomalies via perceptual loss at inference time.

---

Repository structure

Notebook	Purpose
DeepSAM_Pipeline_Main.ipynb	Main pipeline — architecture, training, and AUROC / PRC evaluation
DeepSAM_Segmentation_ViT-based.ipynb	Segmentation comparison: ViT vs Swin-UNet vs U-Net
DeepSAM_SwinUnet_MultiDisease.ipynb	Multi-disease Swin-UNet segmentation + ensemble comparison
DeepSAM_OpenMIBOOD.ipynb	Out-of-distribution detection using OpenMIBOOD benchmark
DeepSAM_MedMNIST_C.ipynb	OOD robustness evaluation on corrupted MedMNIST-C data

---

Notebook descriptions

DeepSAM_Pipeline_Main.ipynb

The main pipeline notebook. Start here to understand and reproduce the core DeepSAM method.

• Architecture definition (convolutional autoencoder)
• Dataset loading (brain tumor MRI, lung disease)
• Model training and evaluation
• AUROC and Precision-Recall Curve (PRC) computation for anomaly detection across multiple disease domains
• Evaluation with negative (healthy) samples

DeepSAM_Segmentation_ViT-based.ipynb

Ablation study comparing ViT-based and classical segmentation architectures used to produce segmentation masks that feed DeepSAM.

• Vision Transformer (ViT) for semantic segmentation
• Swin-UNet (hierarchical Swin Transformer)
• Standard U-Net
• Training loops and Dice-based evaluation for each architecture

DeepSAM_SwinUnet_MultiDisease.ipynb

Trains and evaluates separate Swin-UNet segmentation models for each disease type, then compares ensemble strategies.

• Individual models for: glioma, brain lesion, pituitary, meningioma, COVID-19 pneumonia, pneumonia, normal lung
• Ensemble methods: STAPLE, Feature/Activation Aggregation, average masks
• Head-to-head comparison of ensemble strategies on brain tumor and lung disease datasets

DeepSAM_OpenMIBOOD.ipynb

Evaluates DeepSAM against the OpenMIBOOD out-of-distribution detection benchmark.

• Setup and dataset download for OpenMIBOOD
• Evaluation scripts for Midog, Oasis, and other domains
• Training a new autoencoder with perceptual loss
• Comparison of per-image vs. aggregated perceptual loss scores

DeepSAM_MedMNIST_C.ipynb

Explores robustness to image corruptions using the MedMNIST-C benchmark.

• Loading clean (in-distribution) and corrupted (out-of-distribution) MedMNIST data
• Training the DeepSAM autoencoder on normal samples
• Computing perceptual loss distributions for clean vs. corrupted images
• OOD threshold estimation and detection performance at varying corruption severity levels
• Visualization of score distributions and severity curves

---

Requirements

Dependencies and installation instructions are provided in the ## SETUP section of each notebook. The notebooks were developed in Google Colab (GPU runtime).

Key libraries: PyTorch, torchvision, MedMNIST, medmnistc, scikit-learn, matplotlib.