PaperlayerView exact revision on arXiv

Source-linked AI summary

Does DCGAN-Based Synthetic Augmentation Improve Brain Tumor MRI Classification? An Empirical Study

Irhum Jawad Khan, Talha bin Aslam

arXiv:2609.28508v1eess.IVcs.CV

TL;DR

Medical-image GAN augmentation is increasingly used, but its downstream classification value is uncertain and prior gains may not isolate the effect of synthetic data. This study compares real-only and real+DCGAN training under a fixed Swin Transformer and identical held-out real test set, while measuring FID separately. Accuracy remained 96% in both conditions, with macro F1 unchanged and ROC-AUC slightly lower after augmentation; generated images also showed substantial distributional mismatch.

  • Problem

    Reported classification gains from synthetic medical images do not always isolate their contribution from other training changes, and augmentation benefits can vary across setups.

    Source: §I · p. 1

  • Method

    The study compares a real-only Swin Transformer with the same classifier trained on real images plus 500 class-specific DCGAN images, evaluating both on the same held-out real test set.

    Source: §I · p. 1, C. Classification Model · p. 2

  • Results

    Both conditions achieved 96% accuracy, macro F1 of 0.96, and weighted F1 of 0.96, while ROC-AUC decreased slightly from 0.9954 to 0.9942 after augmentation.

    Source: A. Classification Performance · p. 2

  • Takeaways & Limitations

    Synthetic medical images should be evaluated for distributional and task relevance before their downstream utility is assumed.

    Source: §V · p. 4

    The study is limited to one public dataset, one generator family, one classifier architecture, and one random seed per condition, with no external validation or formal significance test.

    Source: DCGAN-GENERATED IMAGES · p. 3

Abstract

from arXiv · show

Generative adversarial networks (GANs) are increasingly used to augment medical imaging datasets, but synthetic images do not necessarily provide downstream classification benefits. This study investigates whether class-specific Deep Convolutional Generative Adversarial Network (DCGAN) augmentation improves brain tumor classification when the classifier and evaluation set are held constant. Experiments were conducted on 7,200 brain magnetic resonance imaging (MRI) scans across four classes: glioma, meningioma, pituitary tumor, and no tumor. For each class, 1,400 real images were used for training and 400 were reserved for testing. A baseline Swin Transformer classifier was trained using only the real training images and compared with a second model trained using the same real images augmented with 500 DCGAN-generated images per class. Both conditions were evaluated on the identical held-out test set. The two models achieved the same overall accuracy of 96%, while macro F1 remained effectively unchanged and ROC-AUC decreased slightly from 0.987 to 0.982 after augmentation. Class-level analysis showed small redistributions in errors rather than a consistent performance gain. FID values between real and synthetic images ranged from 209.15 to 314.27, indicating substantial distributional differences under the adopted evaluation setup. These results suggest that synthetic augmentation should not be assumed to improve medical image classification and should instead be evaluated for both distributional fidelity and downstream task utility.

I. INTRODUCTION

The study asks whether class-specific DCGAN images improve a fixed brain-MRI classifier under controlled evaluation. Prior work motivates controlled tests because augmentation benefits can vary with the generator, classifier, and synthetic-to-real ratio.

I. INTRODUCTION

  • Medical-imaging datasets can be limited by acquisition cost, privacy constraints, and class imbalance, motivating synthetic augmentation while complicating attribution of reported gains.The introduction notes that some studies do not isolate synthetic images from other training-configuration changes.Source: §I · p. 1

II. RELATED WORK

  • Prior controlled studies indicate that augmentation benefits can depend on the generator, classifier, and synthetic-to-real ratio, [9].Source: §I · p. 1, §II · p. 1

I. INTRODUCTION

  • A matched real-only versus real+DCGAN comparison uses the same held-out real test set and reports no measurable downstream classification gain under the studied conditions.The study also measures per-class FID and qualitatively inspects real and synthetic images.Source: §I · p. 1

II. RELATED WORK

  • FID is used as a standardized distributional diagnostic alongside qualitative inspection, not as a direct measure of clinical quality.Source: §II · pp. 1–2

III. PROPOSED METHODOLOGY

The methodology uses a balanced four-class sample from a public brain-tumor MRI dataset, with a fixed real-image training and held-out testing split for each class.

A. Dataset and Sampling

  • The publicly available Brain Tumor MRI Dataset was sampled across glioma, meningioma, pituitary tumor, and no-tumor classes.Source: A. Dataset and Sampling · p. 2
  • Each class contributed 1,400 real training images and 400 held-out test images, totaling 5,600 training and 1,600 test images.Source: A. Dataset and Sampling · p. 2
  • The same held-out set was used for both classifier conditions, and explicit sample accounting addressed revisions to the hosted dataset.Source: A. Dataset and Sampling · p. 2

B. DCGAN Generation

The study generates class-specific DCGAN samples and evaluates them through a fixed Swin Transformer classification setup.

B. DCGAN Generation

  • A separate DCGAN was trained for each class at 128×128 resolution, with a 100-dimensional latent input and convolutional generator and discriminator blocks.Source: B. DCGAN Generation · p. 2
  • Training used Adam, label smoothing, Gaussian instance noise, and two generator updates per discriminator update; 500 synthetic images were generated per class.The model was trained for 140 epochs.Source: B. DCGAN Generation · p. 2

C. Classification Model

  • The downstream model was a fine-tuned Swin-Tiny checkpoint with a reinitialized four-class head, while test images received only deterministic resizing and normalization.Training augmentation included random crops, flips, rotations, and mild brightness and contrast jitter.Source: C. Classification Model · p. 2

D. Experimental Setup

The experimental setup changes only training-set composition between classifier conditions and evaluates both on the same real test set, with FID measured separately.

D. Experimental Setup

  • The real-only condition used 1,400 images per class, while the real+DCGAN condition added 500 synthetic images per class; classifier training settings were otherwise matched.Both runs stopped at epoch 20.Source: D. Experimental Setup · p. 2

E. Evaluation

  • Classification used the identical 1,600-image real test set and multiple classification metrics, while per-class FID compared 1,400 real training images with 500 corresponding synthetic images.Source: E. Evaluation · p. 2
  • The pipeline trains DCGANs independently by class, holds the real test set constant across classifier conditions, and evaluates FID separately on real versus generated training images.Source: Figure 1 · p. 2

IV. RESULTS AND DISCUSSION

The real-only and real+DCGAN conditions achieved identical accuracy and F1 scores, while ROC-AUC declined slightly after augmentation. Class-level results showed small changes in error patterns rather than a consistent gain.

A. Classification Performance

  • Both conditions achieved 96% accuracy, macro F1 of 0.96, and weighted F1 of 0.96, while ROC-AUC decreased from 0.9954 to 0.9942.Table I summarizes the controlled comparison.Source: A. Classification Performance · p. 2, Table I · p. 2
  • In the controlled comparison, meningioma F1 increased by 0.01 and no-tumor F1 decreased by 0.01, while glioma and pituitary F1 remained unchanged at the class level.Source: A. Classification Performance · p. 2
  • Glioma-to-meningioma errors fell from 39 to 31, while glioma-to-no-tumor errors rose from 18 to 28.These opposing changes left the overall classification outcome unchanged.Source: A. Classification Performance · p. 2

COMPARISON

The comparison presents confusion matrices for the same held-out real test set. The panels distinguish the real-only baseline from the real+DCGAN condition.

  • Fig. 2 presents confusion matrices on the identical held-out real test set.Source: Figure 2 · p. 3
  • The matrices compare classification outcomes for the real-only baseline and real+DCGAN augmentation.Source: COMPARISON · p. 3, Figure 2 · p. 3
  • Panel (a) is the real-only baseline, while panel (b) is the real+DCGAN condition.Source: COMPARISON · p. 3

B. Synthetic Image Quality

FID values were elevated across all four classes, and representative synthetic images showed structural differences from real images. These quality findings are interpreted within a study limited to one dataset, generator family, and classifier setting.

B. Synthetic Image Quality

DCGAN-GENERATED IMAGES

  • The comparison is bounded to one public dataset, one generator family, one classifier architecture, and one random seed per condition.It also lacked external validation and radiologist assessment.Source: DCGAN-GENERATED IMAGES · p. 3

C. Error Analysis

Augmentation redistributed glioma errors without changing its F1 score or producing an overall performance gain. Class-level F1 changes were small and mixed.

C. Error Analysis

  • Glioma-to-meningioma errors decreased from 39 to 31, while glioma-to-no-tumor errors increased from 18 to 28; glioma F1 remained 0.92.Source: C. Error Analysis · pp. 3–4
  • Meningioma F1 increased from 0.95 to 0.96, no-tumor F1 decreased from 0.98 to 0.97, and pituitary tumor F1 remained 0.99.Source: C. Error Analysis · p. 4
  • The errors were redistributed rather than reduced, with no net improvement in overall accuracy or macro-level performance.The single-seed experiment does not establish a causal mechanism for the redistribution.Source: C. Error Analysis · p. 4

D. Comparison with Prior Work

The study’s null result aligns with some prior GAN-augmentation findings, while other reported gains are difficult to compare because they lacked a matched held-out test set. The practical discussion recommends checking distributional fidelity and using controlled evaluation when assessing synthetic-data augmentation.

D. Comparison with Prior Work

  • The null result matches Alrumiah et al.’s finding that GAN-augmented brain MRI did not outperform real-only training, but differs from reported gains in Haque et al. and TumorGANet.The StyleGAN2-ADA study on BRISC 2025 [9] also found augmentation benefits depended on architecture and ratio.Source: D. Comparison with Prior Work · p. 4
  • Neither Haque et al. nor TumorGANet used a matched held-out real test set fixed across conditions, limiting direct comparison with their reported gains.Source: D. Comparison with Prior Work · p. 4

E. Practical Implications

  • Researchers should report FID before incorporating synthetic images and use identical held-out test sets to separate data-composition effects from evaluation-protocol differences.The discussion notes that DCGAN at 128×128 may not reproduce fine anatomical texture and recommends evaluating stronger generators under controlled conditions.Source: E. Practical Implications · p. 4

V. CONCLUSION

In this controlled four-class brain tumor MRI study, adding DCGAN-generated images did not improve accuracy or F1 and slightly reduced ROC-AUC. The generated images also differed substantially from real MRI, supporting evaluation of both distributional and task relevance before assuming downstream utility.

V. CONCLUSION

  • Adding 500 DCGAN-generated images per class did not improve accuracy, macro F1, or weighted F1, and slightly reduced ROC-AUC with the classifier and held-out test set fixed.Source: §V · p. 4
  • Per-class FID values of 209.15–314.27 and qualitative inspection indicated substantial mismatch between generated and real MRI distributions.Source: §V · p. 4
  • The study concludes that synthetic medical images should be assessed for distributional and task relevance before downstream utility is assumed.Future work should compare stronger generators, multiple seeds, and external clinical datasets.Source: §V · p. 4

Related papers

Loading 2609.28508v1…