Source-linked AI summary
Boosted Cascaded Convnets for Multilabel Classification of Thoracic Diseases in Chest Radiographs
Pulkit Kumar, Monika Grewal, Muktabh Mayank Srivastava
TL;DR
ChestX-ray14 is a challenging multilabel dataset with class imbalance and frequent disease co-occurrence. The paper compares losses and develops a boosted cascade that models label dependencies. The cascade outperforms the baseline across all classes and is competitive with published methods.
Problem
ChestX-ray14 has heavy class imbalance and frequent disease co-occurrence, posing challenges for multilabel classification and motivating methods that model label dependencies.
Source: §1 · p. 1Method
The paper compares cross-entropy and PWE losses and proposes a boosted cascade that passes earlier predictions forward and trains later levels on harder examples.
Source: §1 · p. 2, §2.3 · pp. 3–4Results
Boosted cascade models with both losses outperformed the baseline in all classes by a large margin; the best model had better AUCs in 7 of 14 classes than Yao et al. and 1 of 14 than Rajpurkar et al.
Source: §4 · p. 4Takeaways & Limitations
The authors report optimistic results for automatic diagnosis of thoracic diseases and provide comparisons of cross-entropy and pairwise-error losses.
Source: §5 · p. 4The conclusion identifies disease localization and further classification-performance improvement as future work.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Chest X-ray is one of the most accessible medical imaging technique for diagnosis of multiple diseases. With the availability of ChestX-ray14, which is a massive dataset of chest X-ray images and provides annotations for 14 thoracic diseases; it is possible to train Deep Convolutional Neural Networks (DCNN) to build Computer Aided Diagnosis (CAD) systems. In this work, we experiment a set of deep learning models and present a cascaded deep neural network that can diagnose all 14 pathologies better than the baseline and is competitive with other published methods. Our work provides the quantitative results to answer following research questions for the dataset: 1) What loss functions to use for training DCNN from scratch on ChestX-ray14 dataset that demonstrates high class imbalance and label co occurrence? 2) How to use cascading to model label dependency and to improve accuracy of the deep learning model?
1. Introduction
ChestX-ray14 poses a difficult multilabel task because disease classes are imbalanced and frequently co-occur, while chest radiograph interpretation is time-consuming. The paper compares training losses and proposes a boosted cascade to model label dependencies.
1. Introduction
- ChestX-ray14, developed with NLP-derived annotations, provides a large open chest X-ray dataset for automated disease-diagnosis research.The introduction identifies it as the largest open dataset for chest X-ray images at the time described.Source: §1 · p. 1
- Frequent disease co-occurrence and heavy class imbalance make ChestX-ray14 multilabel classification difficult and motivate robust learning that incorporates label dependencies.The dataset contains 112,120 images labeled for 14 diseases; cardiomegaly co-occurs with effusion in 1,060 images.Source: §1 · p. 1
1.1. Related Work
- Earlier multilabel approaches include binary relevance, chain classification, label powersets, recurrent networks, and pairwise-error losses, each with stated limits in modeling or optimization.BR does not account for label interdependence; powerset modeling can become intractable as classes increase, and nonsmooth PWE variants complicate optimization.Source: §1 · pp. 1–2
1. Introduction
- The paper trains DenseNet with BR and compares cross-entropy with PWE loss, then evaluates a cascade for ChestX-ray14 diagnosis.The authors reason that training from scratch on this large dataset could learn features more specific to the diseases than fine-tuning on natural images.Source: §1 · p. 2
- The authors propose a boosted cascade that models complex label dependencies and improves performance over single classifiers trained with cross-entropy or PWE loss.They also train from scratch and compare Binary Relevance (BR) and PairWise Error (PWE) losses.Source: §1 · p. 2
2. Methods
The methods compare binary relevance and pairwise-error losses, then combine cascading with difficulty-weighted training to model label dependencies. Unlike conventional cascades, each succeeding level receives predictions from all preceding levels.
2.1. Binary Relevance
- Binary Relevance treats each class as an independent binary task, using weighted cross-entropy and resampling to address class imbalance.Positive-instance errors are weighted by n/p, where n and p are negative- and positive-instance frequencies.Source: §2.1 · p. 2
2.2. PairWise Error (PWE) loss
- Smooth PWE loss maximizes the margin between positive- and negative-class scores within an example to model inter-class relations.The probability scores are computed from sigmoid outputs, and the BR sampling technique is used to address class imbalance.Source: §2.2 · pp. 2–3
2.3. Boosted Cascade
- Each succeeding cascade level receives predictions from all preceding levels, enabling nonlinear label-dependency modeling and learning from earlier mistakes.The figure caption identifies a four-level cascade architecture.Source: §2.3 · p. 3, Figure 1 · p. 3
- Cascade levels use the entire training set, with later levels weighted toward examples that earlier levels classified with difficulty.The authors contrast this with conventional ensembles trained on separate data splits, which reduce data available to each classifier and limit applicability to neural networks prone to overfitting on smaller datasets.Source: §2.3 · pp. 3–4
3. Implementation
The experiments use ChestX-ray14 with a random 20% test split and DenseNet161-based models. Cascaded predictions average outputs across six levels.
3. Implementation
- The study uses 112,110 frontal ChestX-ray14 images from 30,805 patients, reserving 20% at random for testing.The dataset labels 14 thoracic diseases.Source: §3 · p. 4
- DenseNet161's final fully connected layer is modified to produce two output units per class, reshaped for independent class-versus-background softmax, with weighted cross-entropy.Source: §3 · p. 4
- Both cascaded loss variants use six levels with two fully connected layers per level; inference averages predictions across levels.The levels use ReLU and dropout of 0.5 between fully connected layers.Source: §3 · p. 4
- Training uses SGD with learning rate 0.1 and momentum 0.9, reducing the learning rate to one-tenth after one-third and two-thirds of training.Weights use He norm initialization.Source: §3 · p. 4
4. Results and Discussion
The boosted cascade improved performance over single classifiers for both loss functions and outperformed the dataset baseline across all classes. Its best model had higher AUCs than Yao et al. in 7 of 14 classes and Rajpurkar et al. in 1 of 14, with comparable AUCs in the remaining classes.
4. Results and Discussion
- The boosted cascade with either loss outperformed the dataset baseline in all classes by a large margin.The BR cross-entropy and PWE baselines outperformed it in 9 and 10 classes, respectively.Source: §4 · p. 4
- BR with cross-entropy and PWE loss performed comparably, while adding the boosted cascade increased performance with both losses.Table [1] reports ROC AUCs for binary classification across all classes.Source: §4 · p. 4
- The best model had better AUCs than Yao et al. [11] in 7 of 14 classes and Rajpurkar et al. [6] in 1 of 14, with comparable AUCs in the rest.The comparison is reported for the ChestX-ray14 dataset.Source: §4 · p. 4
5. Conclusion
The paper compares cross-entropy and pairwise-error losses for multilabel classification on ChestX-ray14 and reports results comparable to the state of the art. Its cascade improves deep-learning performance while modeling label dependencies; future work includes disease localization and further classification improvement.
5. Conclusion
- The experiments compare cross-entropy and pairwise-error losses for multilabel classification on ChestX-ray14.The work reports results comparable to the state of the art.Source: §5 · p. 4
- The cascade improves deep-learning model performance while modeling label dependencies.Source: §5 · p. 4
- The authors suggest disease localization and improved classification performance as future work.Source: §5 · p. 4