Design and Implement a Flexible, Lightweight Deep Learning Model for Restoring and Enhancing Leukemia Blood-Smear Images.

Abstract

Diagnosis of leukemia can be difficult using peripheral blood smears because the smears can have artifacts that pathologists may not be able to recognize, such as sensor noise, optical blur, uneven illumination, and low contrast that can obscure the fine nuclear and cytoplasmic details they use to make a diagnosis. This paper proposes a lightweight convolutional network called LARNet (Lightweight Adaptive Restoration Network), which consists of a residual encoder–decoder network that integrates channel and spatial attention mechanisms, and is trained with a hybrid loss function combining pixel fidelity loss, structural similarity (SSIM) loss, and edge-preservation loss. With 147,007 trainable parameters (0.56 MB in single precision), it is easily deployable in resource-limited systems, and it learns a correction for the input, instead of a filter response. The model was trained for 60 epochs on the C-NMC 2019 blood-smear corpus, and achieved the best PSNR of 24.82 dB and the SSIM of 0.630 for the validation set. LARNet had a mean PSNR of 25.32 ± 2.38 dB on the independent test comparison, which was a gain of 5.2 dB compared to the best classical baseline tested (Non-Local Means, 20.08 dB), and a considerable improvement over Median Filtering, Gaussian Filtering and CLAHE. This pixel fidelity gain is substantial, uniform and lies far outside the range of the classical baselines and is the main, headline finding of this paper. Although this is below the raw SSIM score for Non-Local Means (0.762), LARNet also shows significant improvement over the unrestored degraded image (0.602) and CLAHE (0.621); and the smaller SSIM difference between images of LARNet (2.38 dB) compared with the classical methods (3.88–5.01 dB), suggests a more stable response to degradation across images. Also, beyond denoising, LARNet can be extended to solve a more challenging task, namely, restore and enhance an image simultaneously (restoration + enhancement), which is the dual objective that motivates this work, through an explicit restoration+enhancement pipeline (contrast, sharpness, and stain-colour normalisation). Combined, the results render LARNet as a compact, input adaptive restoration model, which achieves a significant and statistically significant improvement of fidelity over the classical restoration models, for leukemia blood-smear imagery, while the structural-similarity/latency trade-off is also clearly reported in this paper, so that the actual contribution can be assessed.