Source-linked AI summary
Anatomy-specific classification of medical images using deep convolutional nets
Holger R. Roth, Christopher T. Lee, Hoo-Chang Shin, Ari Seff, Lauren Kim, Jianhua Yao, Le Lu, Ronald M. Summers
TL;DR
Automated anatomy classification is difficult because human anatomy varies substantially, yet it is important for computer-aided diagnosis. This paper trains a ConvNet on PACS-mined, DICOM-labeled axial CT key-images and enriches training data with plausible spatial deformations. With augmentation, testing classification error falls to 5.9% and average AUC reaches 0.998; the classifier can initialize further analysis, although lower-abdomen slices remain a scope boundary.
Problem
The paper addresses reliable automated classification of human anatomy, an important prerequisite for computer-aided diagnosis.
Method
A supervised ConvNet learns five anatomical classes from PACS-mined axial CT key-images labeled using radiological reports and DICOM tags, with spatial data augmentation.
Results
5.9% testing classification error and average AUC 0.998 are achieved using ConvNets and data augmentation.
Takeaways & Limitations
The anatomy classifier could initialize further disease- and organ-specific computer-aided detection or diagnosis analysis.
Takeaways & Limitations
Some lower-abdomen slices produce erroneously high lung or legs probabilities, motivating a possible additional lower-abdomen class.
Abstract
from arXiv · showhide
Automated classification of human anatomy is an important prerequisite for many computer-aided diagnosis systems. The spatial complexity and variability of anatomy throughout the human body makes classification difficult. "Deep learning" methods such as convolutional networks (ConvNets) outperform other state-of-the-art methods in image classification tasks. In this work, we present a method for organ- or body-part-specific anatomical classification of medical images acquired using computed tomography (CT) with ConvNets. We train a ConvNet, using 4,298 separate axial 2D key-images to learn 5 anatomical classes. Key-images were mined from a hospital PACS archive, using a set of 1,675 patients. We show that a data augmentation approach can help to enrich the data set and improve classification performance. Using ConvNets and data augmentation, we achieve anatomy-specific classification error of 5.9 % and area-under-the-curve (AUC) values of an average of 0.998 in testing. We demonstrate that deep learning can be used to train very reliable and accurate classifiers that could initialize further computer-aided diagnosis.
1. INTRODUCTION
The paper targets automated classification of axial CT images into five anatomical classes to support computer-aided detection and diagnosis. It establishes training and testing ground truth by mining radiological reports and associated DICOM tags for key-images.
- Automated anatomy classification is an important component of many computer-aided detection and diagnosis systems.
- The study classifies axial CT images into five anatomical classes: neck, lungs, liver, pelvis and legs.
- Ground truth is established by mining radiological reports that refer to key-images and associated DICOM image tags.
- The authors report that computer vision and medical image computing techniques replicate these classes with low error rates.
2. METHOD
The method trains a supervised five-layer ConvNet on PACS-derived, DICOM-labeled CT key-images, using learned filters, pooling, and plausible spatial deformations to enrich the data.
- Convolutional networks: ConvNets are used to build an anatomy-specific CT classifier with five cascaded layers of supervised convolutional filters.
- Convolutional networks: Max-pooling summarizes neighboring feature responses, while a fully connected network and softmax layer produce anatomy-specific class probabilities.
- Data mining of key-images: Medical images are retrieved from a hospital PACS archive and labeled using StudyDescription and BodyPartExamined DICOM tags, with manual correction when necessary.
- Data mining of key-images: Images spanning multiple anatomical classes are duplicated and assigned separate labels, with roughly equal probabilities assigned in transition regions.
- Data augmentation: Data augmentation applies random translations, rotations, and thin-plate-spline non-rigid deformations to generate plausible physical variations.
- Data augmentation: Augmented images are rescaled to 256 × 256 and used to train a standard multi-class ConvNet architecture.
3. RESULTS
Data augmentation improved anatomy-specific ConvNet classification on the original test images, while t-SNE showed separation of most anatomical classes and qualitative testing captured lung–liver interfaces.
- Quantitative performance: 5.9% testing classification error and 0.998 average AUC followed data augmentation, improving from 9.6% and 0.994, respectively.Confusion matrices showed fewer misclassifications on the original test set after augmentation.
- Feature-space analysis: Most anatomical classes formed clearly separated groups in the ConvNet t-SNE feature space.The embedding used a subset of test key-images, with colors indicating ground-truth labels.
- Feature-space analysis: An overlapping feature cluster appeared at the lung–liver interface, where key-images contain both anatomies.The overlap reflects mixed anatomy rather than an isolated classification error.
- Volume evaluation: A full torso CT examination was classified slice by slice using anatomy-specific probability profiles.The qualitative evaluation applied the trained classifier to a volume with dimensions [512, 512, 652] and [0.98, 0.98, 1.5] mm voxel spacing.
- Volume evaluation: The classifier assigned roughly equal probabilities at the diaphragm’s lung–liver interface and processed the examination in less than 1 minute.This result was obtained on a modern desktop computer with a GPU card.
4. DISCUSSION
The discussion positions manually labeled key-images as a practical basis for anatomy-specific classification despite loose report–image associations. It also identifies lower-abdomen errors and suggests extensions for broader anatomy modeling and downstream CADx or CADe analysis.
- Interpretation and related work: Manually labeled key-images provide the training basis, whereas associations between clinical reports and image scans can be too loose for reliable retrieval.The paper therefore focuses on anatomy-specific classification from manually labeled images rather than report-based retrieval.
- Interpretation and related work: The ConvNet modeled images rescaled to 256 × 256 and generalized to large variations in medical imaging data with promising quantitative and qualitative results.This resolution was closer to the original 512 × 512 images than more aggressively reduced inputs.
- Limitations and extensions: Erroneously high lung or leg probabilities occurred for some lower-abdominal slices, motivating an additional lower-abdomen class.The paper also suggests image-scale augmentation to model patient-size variation.
- Practical implication: An anatomy classifier could initialize more detailed disease- and organ-specific computer-aided detection or diagnosis.This proposed use is presented as a downstream application of the anatomy classification system.