Evaluating JPEG2000 compression effects on prediction accuracy and storage efficiency in lung cancer whole-slide images: a deep learning-based analysis

Whole-slide images (WSIs) in digital pathology are generated at gigapixel resolution, leading to substantial storage and data management demands. Lossy image compression is commonly used to mitigate these challenges. However, its impact on deep learning (DL)-based image analysis, particularly for detailed segmentation tasks, remains insufficiently characterized. DL models may be sensitive to subtle compression artifacts, underscoring the need for systematic evaluation.

We evaluated JPEG2000 compression effects on tile-level tumor segmentation using lung adenocarcinoma WSIs. Eighty-six WSIs annotated by four board-certified pathologists were subdivided into nonoverlapping 512 × 512 tiles at 20× magnification. To assess architectural robustness, we trained CNN-based models (U-Net with MobileNetV3, ResNet-18, ResNet-50 encoders) and a Transformer-based model (SegFormer) exclusively on lossless PNG images. These models were evaluated on test tiles compressed using JPEG2000 at rates ranging from 1 (lossless) to 80. Segmentation performance was assessed using the Dice coefficient. Image quality was quantified using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and the no-reference Perception-based Image Quality Evaluator (PIQE). Storage efficiency was evaluated relative to uncompressed TIFF images.

Segmentation performance remained stable at mild compression levels, with a marginal decrease of approximately 1.7% observed in the most robust model (MobileNetV3) up to JPEG2000 compression rates of 10 compared with the lossless baseline. In this range, image quality metrics remained high (PSNR ≥ 33 dB, SSIM ≥ 0.97), while storage requirements were reduced by approximately 90%. At higher compression rates (≥ 15), segmentation performance declined progressively, accompanied by marked degradation in image quality and increased perceptual artifacts. Comparative analysis across model architectures revealed that convolutional neural network (CNN)-based models exhibited greater robustness to compression artifacts than Transformer-based models, which demonstrated an earlier and steeper performance decline. Analysis of model confidence maps further showed that boundary uncertainty increased with compression severity, even before substantial Dice score degradation became apparent.

JPEG2000 compression at mild rates preserves DL-based segmentation performance while substantially reducing storage requirements in digital pathology workflows. These findings provide quantitative evidence on compression-performance trade-offs for tile-level image analysis under controlled experimental conditions and highlight architecture-dependent sensitivity to compression artifacts.