Enhancing MRI Brain Tumor Edge Detection: A Hybrid Preprocessing Approach Utilizing CLAHE

1 Introduction

• Manual tumor-boundary delineation is subjective, time-consuming, and marked by substantial disagreement among expert raters.The section notes that this bottleneck can delay surgical interventions in high-throughput clinical settings.Source: §1 · p. 2
• Deep-learning segmentation can require many annotated samples and high-end GPUs, while its black-box behavior limits interpretability in clinical decisions.These drawbacks motivate continued interest in mathematically interpretable classical approaches.Source: §1 · p. 2
• MRI noise, intensity bias, and Rician-distributed magnitude noise can destabilize Canny’s fixed smoothing and hysteresis parameters.The section also notes that sequential parameter sweeps have been used to automate threshold selection, often with global morphological preprocessing.Source: §1 · p. 2
• Global preprocessing can suppress weak gradients at diffuse tumor–edema boundaries, motivating localized CLAHE enhancement to improve boundary recovery.The proposed hypothesis is that adding optimally configured CLAHE before the automated parameter sweep will improve edge-detection performance.Source: §1 · pp. 2–3
• The proposed pipeline adds an optimally configured CLAHE layer to automated edge detection, prioritizing stability and real-time execution over computationally heavy metaheuristic optimization.The introduction frames the approach as a practical operational model for neuro-oncology.Source: §1 · p. 3
• The study compares standard hybrid-stretching pipelines with the proposed hybrid model on the Cheng et al. (2016) benchmark using SSIM, PSNR, F1-Score, and Recall.The introduction says the analysis examines clinical precision–recall trade-offs.Source: §1 · p. 3

2.1 Deep Learning vs. Classical Segmentation

• Deep-learning segmentation offers strong accuracy but depends on large annotated datasets and substantial computation, while classical and unsupervised approaches remain alternatives with their own MRI intensity-inhomogeneity challenges.The section also describes Canny as producing single-pixel edges, unlike noise-sensitive Sobel and Prewitt operators.Source: §2.1 · pp. 3–4, §2.2 · p. 4

2.2 Advancements in Automated Edge Detection

• Automated Canny tuning includes similarity-based sweeps, genetic and Otsu-based methods, and metaheuristic optimization, whose search overhead can conflict with real-time deployment.The section also discusses learned edge detectors, which inherit annotation and inference-cost burdens.Source: §2.2 · pp. 4–5

2.3 Contrast Enhancement and CLAHE

• Global equalization can amplify background noise, whereas CLAHE clips histogram amplification to boost local contrast while constraining noise.Brain-tumor studies combining enhancement with segmentation or classification motivate enhancement before edge detection.Source: §2.3 · p. 5

2.4 The Canny Edge Detection Framework

• Canny’s sequence smooths the image with a Gaussian kernel, calculates gradient magnitude and direction, thins edges by non-maximum suppression, then applies two-threshold hysteresis.Weak pixels are retained only when connected to strong-edge pixels.Source: §2.4 · p. 5

3.1 Base Morphological Preprocessing (Phase 1)

• Both experimental arms resize grayscale MRI slices to 256 × 256, apply top-hat filtering and a Gaussian-likelihood hybrid stretch, then renormalize intensities to [0, 1].Source: §3.1 · p. 6

3.2 The Optimally Configured CLAHE Layer - Proposed Addition (Phase 2)

• CLAHE divides images into non-overlapping contextual tiles, equalizes histograms locally, and uses fixed MRI settings (num_tile 8, 32 × 32-pixel grids; clip_limit 0.01), avoiding per-image metaheuristics’ 65x time penalty and ground-truth overfitting risk.Source: §3.2 · p. 8

3.3 Automated Sequential Parameter Tuning (Phase 3)

• A deterministic sequential coordinate-wise sweep tunes all three Canny parameters, locking each at the value yielding maximum consecutive edge-map pixel-wise agreement.Source: §3.3 · p. 8

3.4 Dataset Configuration

• The curated dataset contains 256 × 256 grayscale brain MRI images, primarily T1-weighted and FLAIR, with varied tumor pathology and scanner noise.Source: §3.4 · p. 8

3.5 Evaluation Metrics

• Evaluation compares the pipelines against a ground-truth edge map using quantitative measures that include recall, structural similarity, distortion, and execution time.Source: §3.5 · pp. 8–9

• The benchmark was run across all MRI images from the Cheng et al. (2016) dataset.Source: §4 · p. 9
• Table 1 reports the mean and standard deviation for each evaluation metric across the two pipelines.Source: §4 · p. 9
• The reported results are organized by metric to compare the baseline and proposed pipelines.Source: §4 · p. 9

5.1 Quantitative Interpretation and Clinical Relevance

• Recall rose from 0.3327 with Vanilla PST-CED to 0.4636 with Fixed CLAHE, a reported relative improvement of nearly 39%.The comparison is reported for the proposed preprocessing change; Table 1 is identified as the baseline-versus-proposed performance comparison over MRI images.Source: §5.1 · pp. 9, 11, Table 1 · p. 11
• Precision declined from 0.5771 with Vanilla PST-CED to 0.5332 with the proposed method, with slight drops in Accuracy and PSNR.The authors say localized contrast stretching can also amplify MRI noise or minor healthy-tissue boundaries, increasing false-positive edge pixels.Source: §5.1 · p. 11
• F1-Score increased from 0.4161 to 0.4887 and Kappa Index from 0.3805 to 0.4477 with the proposed method.The authors attribute these gains to Recall improvements outweighing Precision losses.Source: §5.1 · p. 11, Table 1 · p. 11
• SSIM improved from 0.6421 to 0.6661, which the authors interpret as better preservation of perceptual structural integrity and spatial coherence.The interpretation is reported despite a denser edge map.Source: §5.1 · p. 11
• Execution time increased by 0.06 seconds, from 0.80s to 0.86s per image, with CLAHE preprocessing.The authors characterize this as less than a tenth of a second of overhead alongside the reported sensitivity gain.Source: §5.1 · p. 11

6 Conclusion

• Adding fixed-parameter CLAHE to PST-CED improved detection of subtle tumor boundaries compared with relying on global morphological preprocessing alone.The authors describe this as a structural extension of the MRI edge-detection framework.Source: §6 · p. 12
• The proposed pipeline is reported to improve Recall by 39% and elevate F1-Score and SSIM, while introducing a slight increase in minor false edges.The authors also describe a negligible drop in precision.Source: §6 · p. 12
• The pipeline averaged 0.86 seconds per image, which the authors characterize as a virtually zero computational penalty and a real-time operational model.The conclusion presents it as a practical diagnostic tool for clinical neuro-oncology.Source: §6 · p. 12