Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Active Learning with Variational Autoencoder for Radiology Data Generation in Veterinary Medicine

In-Gyu Lee, Jun-Young Oh, Hee-Jung Yu, Jae-Hwan Kim, Ki-Dong Eom, Ji-Hoon Jeong

TL;DR

This work tackles the scarcity of veterinary radiology data for computer-aided diagnosis by introducing a generative active learning framework that pairs a ProjectedGAN-based data generator with a VAE-based data-filtering query. The data generation phase continuously improves synthetic radiographs, while the query phase selects high-latent-space-similarity images to augment the training set. Evaluation shows decreasing Fréchet Inception Distance and improved classification metrics for cardiomegaly radiographs, with Cycle-3 often delivering peak performance. Limitations include reliance on a GAN variant and radiography-only experiments; future work points toward diffusion models and additional imaging modalities to further enhance CAD in veterinary medicine.

Abstract

Recently, with increasing interest in pet healthcare, the demand for computer-aided diagnosis (CAD) systems in veterinary medicine has increased. The development of veterinary CAD has stagnated due to a lack of sufficient radiology data. To overcome the challenge, we propose a generative active learning framework based on a variational autoencoder. This approach aims to alleviate the scarcity of reliable data for CAD systems in veterinary medicine. This study utilizes datasets comprising cardiomegaly radiograph data. After removing annotations and standardizing images, we employed a framework for data augmentation, which consists of a data generation phase and a query phase for filtering the generated data. The experimental results revealed that as the data generated through this framework was added to the training data of the generative model, the frechet inception distance consistently decreased from 84.14 to 50.75 on the radiograph. Subsequently, when the generated data were incorporated into the training of the classification model, the false positive of the confusion matrix also improved from 0.16 to 0.66 on the radiograph. The proposed framework has the potential to address the challenges of data scarcity in medical CAD, contributing to its advancement.

Generative Active Learning with Variational Autoencoder for Radiology Data Generation in Veterinary Medicine

TL;DR

This work tackles the scarcity of veterinary radiology data for computer-aided diagnosis by introducing a generative active learning framework that pairs a ProjectedGAN-based data generator with a VAE-based data-filtering query. The data generation phase continuously improves synthetic radiographs, while the query phase selects high-latent-space-similarity images to augment the training set. Evaluation shows decreasing Fréchet Inception Distance and improved classification metrics for cardiomegaly radiographs, with Cycle-3 often delivering peak performance. Limitations include reliance on a GAN variant and radiography-only experiments; future work points toward diffusion models and additional imaging modalities to further enhance CAD in veterinary medicine.

Abstract

Recently, with increasing interest in pet healthcare, the demand for computer-aided diagnosis (CAD) systems in veterinary medicine has increased. The development of veterinary CAD has stagnated due to a lack of sufficient radiology data. To overcome the challenge, we propose a generative active learning framework based on a variational autoencoder. This approach aims to alleviate the scarcity of reliable data for CAD systems in veterinary medicine. This study utilizes datasets comprising cardiomegaly radiograph data. After removing annotations and standardizing images, we employed a framework for data augmentation, which consists of a data generation phase and a query phase for filtering the generated data. The experimental results revealed that as the data generated through this framework was added to the training data of the generative model, the frechet inception distance consistently decreased from 84.14 to 50.75 on the radiograph. Subsequently, when the generated data were incorporated into the training of the classification model, the false positive of the confusion matrix also improved from 0.16 to 0.66 on the radiograph. The proposed framework has the potential to address the challenges of data scarcity in medical CAD, contributing to its advancement.
Paper Structure (12 sections, 6 equations, 4 figures, 1 table, 1 algorithm)

This paper contains 12 sections, 6 equations, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Generative Active Learning with VAE
  3. Dataset
  4. Proposed Framework
  5. Data generating phase
  6. Query phase
  7. Classification phase
  8. Experiments
  9. Data generating phase
  10. Classification phase
  11. Discussion
  12. Conclusion and Future works

Figures (4)

  • Figure 1: The data preprocessing pipeline for training a generative model. If doctors draw annotations for diagnosis, the annotations are extracted to create masks. Then, image inpainting techniques are applied to remove the annotations. Subsequently, the resolution of the images is standardized.
  • Figure 2: Overall flow of the proposed framework. (a) The projectedGAN is trained with filtered image data and real image data to generate a new radiographic image. (d) VAE trained with 100 original data are used to filter generated images using a query strategy. (c) The top 10% cosine similarity of the data is added to the training dataset. (b) Finally, classification is performed using the object detection model after labeling to prove the usefulness of the data.
  • Figure 3: Generation results of each cycle. The following is data generated by learning cardiomegaly data.
  • Figure 4: Confusion matrix results of classification phase. The above results represent training and testing results using cardiomegaly data. Among the five sessions, the confusion matrix of the session with the highest accuracy is presented more prominently.