Please enable JavaScript.
Coggle requires JavaScript to display documents.
Survey - Coggle Diagram
Survey
Explainability and Interpretability
Taxonomy: intrinsic vs post-hoc, global vs local, model-specific vs model-agnostic
Explainability in VLMs: language as an explanation, grounding, report-level rationales, faithfulness of generated text
Why explainability matters in medicine (trust, safety, regulation, bias detection)
Limitations and risks
Explainability applied to CT models
Methods: Attribution/saliency (Grad-CAM, Integrated Gradients, LRP, SHAP, LIME), Attention-based explanations for transformers , Concept-based explanations (TCAV, concept bottleneck models) , Counterfactual and example-based explanations
Vision-Language Models
Architectures: dual-encoder, fusion-encoder, and LLM-based multimodal models (BLIP, Flamingo, LLaVA, GPT)
Medical VLMs: BiomedCLIP, MedCLIP, LLaVA-Med, Med-PaLM M, RadFM
Foundations: multimodal learning, contrastive learning (CLIP), image-text alignment
VLMs for CT specifically: 3D CT-text alignment (e.g., CT-CLIP, M3D), radiology report generation, visual question answering, zero-shot diagnosis
Challenges: hallucination, 3D volume handling, data scarcity, evaluation metrics
Deep Learning for CT Scan Analysis
Applications by anatomy and disease ( chest/lung , abdomen , brain , others)
Preprocessing and data handling (resampling, normalization, augmentation, patch-based training, Low-dose CT denoising)
CT fundamentals relevant to AI (Hounsfield units, windowing, slice thickness, 2D vs 3D data)
Limitations of current CT-based approaches
Deep Learning for Medical Imaging
Medical-specific architectures (U-Net variants, nnU-Net, TransUNet, Swin UNETR)
Imaging modalities overview (X-ray, MRI, CT, ultrasound, PET)
Datasets and benchmarks (e.g., MedMNIST, BraTS, LIDC-IDRI, CheXpert, MIMIC-CXR)
Domain-specific challenges: limited annotations, class imbalance, privacy
Deep Learning : Foundations
Training paradigms
self-supervised
semi-supervised
supervised
transfer learning
Optimization, regularization, and common challenges (overfitting, data scarcity, domain shift)
Core architectures
RNNs/LSTMs
CNNs
MLPs
Autoencoders and generative models
Transformers and attention mechanisms
Deep Learning for Computer Vision
Semantic and instance segmentation (FCN, U-Net, Mask R-CNN, SAM)
Object detection (R-CNN family, YOLO, DETR)
Image classification (AlexNet to ResNet, EfficientNet, ViT, hybrid CNN-Transformer models)