Adam Framework for Medical Image Foundation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing self-supervised learning methods lack the capability to appreciate the foundation of medical imaging-human anatomy, leading to incompetence in generating effective foundation models for medical image classification and segmentation.
Innovation Solution
A novel training strategy using hierarchical self-supervised contrastive learning, specifically the Adam framework, which decomposes and perceives anatomy progressively in a coarse-to-fine manner through Gray-coded image division, enabling the model to learn dense and semantics-rich anatomical embeddings that preserve locality and compositionality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing self-supervised learning methods are used, then model training can proceed without labeled data, but the model fails to capture anatomical structures and patterns
Solution Approach 1:
The patent segments the learning process into multiple stages: first learning global anatomical structures, then progressively learning finer-grained anatomical details. This hierarchical segmentation allows the model to capture anatomical patterns at different levels of detail, resolving the contradiction between using unlabeled data and capturing anatomical information.
Solution Approach 2:
The patent applies preliminary action by first pre-training the model on unlabeled medical images to learn general anatomical structures, then fine-tuning with labeled data for specific tasks. This preliminary learning of anatomical patterns without task-specific labels enables the model to retain anatomical understanding while adapting to specific medical tasks.
2Loss of information
If conventional supervised learning is used, then the model can learn accurate anatomical patterns, but it requires large amounts of labeled data which are scarce in medical imaging
Solution Approach 1:
The patent implements self-service through self-supervised learning mechanisms where the model learns from unlabeled medical images by creating its own supervision signals. The model identifies anatomical structures and patterns autonomously without requiring human-annotated labels, thereby reducing dependency on scarce labeled data while maintaining anatomical understanding.
Solution Approach 2:
The model performs preliminary learning on unlabeled data to acquire general anatomical knowledge before task-specific fine-tuning. This preliminary action enables the model to build a robust anatomical foundation without requiring labeled data, reducing the quantity of labeled data needed for subsequent task-specific learning.
3Adaptability or versatility
If the model learns general patterns, then it can adapt to multiple medical tasks, but it lacks specialized knowledge for specific anatomical structures
Solution Approach 1:
The patent segments anatomical learning into hierarchical levels, where the model first learns global anatomical structures and then progressively learns finer-grained details. This segmentation allows the model to maintain both general adaptability across tasks and specialized precision for specific anatomical structures at different granularities.
Solution Approach 2:
The patent employs dynamic adaptation through fine-tuning mechanisms that allow the model to specialize its general anatomical knowledge for specific tasks. The model dynamically adjusts its parameters during task-specific fine-tuning to achieve both versatility across multiple medical tasks and precision for specific anatomical structures.
Data Source
AI summary
Systems, methods, and apparatuses for learning foundation models from anatomy in medical imaging for use with medical image classification and/or image segmentation in the context of medical image analysis. Exemplary systems include means for receiving medical images; extracting human anatomical patterns from the medical images; generating a foundation model via learning the human anatomical patterns from within the medical images received, resulting in generic representations of the human anatomical patterns; wherein the learning includes: first learning prominent objects from within the medical images received corresponding to the human anatomical patterns; and secondly learning detailed parts within the learned prominent objects corresponding to sub-portions of the generic representations of the human anatomical patterns; wherein the learning further includes executing a self-supervised contrastive learning framework, including: executing an anatomy decomposer (AD) of the self-supervised contrastive learning framework which guides the generated foundation model to conserve hierarchical relationships of anatomical structures within the medical images received; and executing a purposive pruner (PP) of the self-supervised contrastive learning framework which forces the model to capture more distinct representations for different anatomical structures at varying granularity levels; and outputting the generated foundation model for use in processing medical images which form no part of the medical images received and used for training the generated foundation model.