TL;DR
This paper introduces Unsupervised Body Part Regression (UBR), a novel deep learning method that learns to predict the normalized position of anatomical body parts within CT slices without requiring any labeled data. By leveraging the inherent superior-inferior ordering of CT slices, UBR offers a fast, accurate, and annotation-free approach for medical image analysis, with demonstrated utility in network initialization and anomaly detection.
What Problem Does This Paper Address?
Accurately identifying body parts in medical images, particularly CT scans, is crucial for various applications like automated scan planning and initializing computer-aided detection (CAD) systems. While deep learning has shown great promise, it typically demands vast amounts of meticulously labeled training data, which is expensive and time-consuming to acquire. Existing unsupervised methods often still require some form of fine-grained labels for effective fine-tuning. Furthermore, many approaches discretize the body into distinct parts, which can be problematic at transition regions and lack the granularity to differentiate slices within the same part. This paper aims to bridge this gap by developing an unsupervised method that can learn a continuous representation of body part location from unlabeled CT volumes, exploiting the natural ordering of slices.
Methods
The core of this work is the Unsupervised Body Part Regression (UBR) algorithm, which uses a Convolutional Neural Network (CNN). UBR's novelty lies in its unsupervised training strategy and loss functions. Instead of relying on explicit labels, UBR learns by exploiting the superior-inferior ordering of CT slices within a volume.
The training process involves randomly selecting groups of equidistant slices from unlabeled CT volumes. The CNN architecture is based on established models like VGG-16, with some modifications. The key innovation is the loss function, which comprises two terms:
- Order Loss ($L_{order}$): This logistic loss encourages slices with higher indices (further down the superior-inferior direction) to have higher predicted scores. This enforces the qualitative ordering of body parts.
- Distance Loss ($L_{dist}$): This loss function, using smooth L1 loss, constrains the predicted scores to increase linearly with slice index. It ensures that the numeric difference between scores is proportional to the spatial distance between slices, making the regression more precise.
The total loss is a combination of these two terms ($L = L_{order} + L_{dist}$). This "self-organization" process allows the network to learn body part knowledge from inter-slice relationships without any manual annotations.
Datasets
The paper uses two main datasets:
- Training Dataset: 800 random, unlabeled CT volumes from 420 subjects, with scan ranges varying (chest, abdomen, pelvis). These are sourced from the hospital's PACS.
- Testing Dataset: 18,195 CT slices randomly sampled from 260 CT volumes of 140 new subjects. For quantitative evaluation, these slices were manually labeled into three common classes: chest, abdomen, and pelvis.
For the application experiments, the DeepLesion dataset [10] was used, which contains 32,120 CT slices with bounding-box annotations for various lesions.
Key Insights and Ideas
- Leveraging Intrinsic Data Structure: The paper demonstrates that the inherent superior-inferior ordering of CT slices is a powerful signal for unsupervised learning, eliminating the need for explicit anatomical labels.
- Continuous Regression for Granularity: Moving beyond discrete body part classification, UBR's continuous score output provides a fine-grained representation, useful for identifying subtle positional information and transitions.
- Self-Organization for Learning: The "self-organization" metaphor highlights how local inter-slice relationships, guided by the novel loss functions, lead to a global understanding of body part positioning.
- Annotation-Free Power: The ability to train effectively on completely unlabeled data is a significant advantage in medical imaging, where labeling is a major bottleneck.
- Versatile Feature Extraction: The learned features from UBR are shown to be useful not only for body part regression but also for initializing networks for other tasks like lesion detection.
Key Results
- Body Part Classification Accuracy: When using UBR's regressed scores to classify slices into chest, abdomen, and pelvis, UBR (with only 2 labeled volumes for threshold calibration) achieved 95.99% accuracy. This outperforms a prior self-supervised method [3] that required fine-tuning with 88 labeled volumes (95.28%).
- Feature Performance: Features extracted from UBR's pool6 layer, when used to train a logistic regressor on 88 labeled volumes, yielded an accuracy of 98.41%, approaching supervised methods.
- Application 1: Network Initialization: In lesion detection on the DeepLesion dataset, initializing a Faster RCNN model with UBR features (DeepLesion-UBR) significantly improved performance over random initialization. The doubly fine-tuned DeepLesion-ImageNet+UBR scheme achieved the highest accuracy (74.82%).
- Application 2: Anomaly Detection: UBR can identify anomalies by detecting deviations from the expected linear trajectory of slice scores in normal volumes, flagging issues like ascitic fluid or atypical scan directions.
- Speed: The inference time per slice is a rapid 4 ms.
Limitations and Future Work
- Supervised Baseline: The paper acknowledges that a fully supervised method with extensive labeling (98.84% accuracy) still achieves the highest classification performance. However, UBR's advantage lies in its lack of annotation requirements.
- ImageNet Pre-training: While UBR performs well, the combination with ImageNet pre-training (DeepLesion-ImageNet+UBR) yielded slightly better results for lesion detection, suggesting that ImageNet's broad visual features are still highly beneficial. The authors suggest UBR can further "familiarize" ImageNet models with CT image statistics.
- Parameter Sensitivity: The paper notes that the choice of the number of slices per group ($m$) can impact accuracy, and the distance loss might become overly strict with very large $m$.
- Future Work: The authors demonstrate UBR's utility in network initialization and anomaly detection, implying these are promising avenues for further exploration. The method's adaptability to other body parts and imaging modalities could also be explored.
Summary updated