PaperlayerView exact revision on arXiv

Source-linked AI summary

Contrast Enhancement or Noise Reduction? On Improving Cervical Cancer Classification

Ach Khozaimi, Ulfatun Nahdhiyah

arXiv:2610.11086v1cs.CV

TL;DR

Pap smear image classification research has paid limited attention to preprocessing, leaving the relative benefits of contrast enhancement and noise reduction unclear. This study compares CLAHE and PMD with three CNN architectures on SIPaKMeD. CLAHE consistently outperforms PMD and original images in accuracy.

  • Problem

    The impact of image preprocessing on CNN-based cervical cancer classification remains insufficiently explored, and the relative benefits of contrast enhancement versus noise reduction are unclear.

    Source: INTRODUCTION · p. 2

  • Method

    The study evaluates CLAHE and PMD preprocessing with ResNet-34, DenseNet-121, and MobileNet-V2 on the SIPaKMeD dataset using a confusion matrix.

    Source: INTRODUCTION · p. 2

  • Results

    CLAHE outperformed PMD and original images, reaching accuracies of 84.16% for ResNet-34, 83.17% for DenseNet-121, and 84.65% for MobileNet-V2.

    Source: CONCLUSION · p. 10

  • Takeaways & Limitations

    Contrast enhancement was more effective than noise reduction for improving CNN accuracy in cervical cancer classification on SIPaKMeD.

    Source: CONCLUSION · p. 10

Abstract

from arXiv · show

Purpose: Cervical cancer is one of the leading causes of mortality worldwide. Deep learning has shown promising performance in medical image classification. The influence of image preprocessing algorithms on classification performance remains insufficiently investigated in the literature. This research aims to evaluate the impact of image preprocessing algorithms on the performance of CNNs for Pap smear image classification. Methods: Three CNN architectures (ResNet-34, MobileNet-V2, and DenseNet-121) were trained and evaluated using the SIPaKMeD dataset. Two preprocessing algorithms were applied: the PMD filter for noise reduction and CLAHE for contrast enhancement. The model performance was assessed using a confusion matrix. Results: Preprocessing improved the classification performance of all models. CLAHE significantly increased the accuracy of ResNet-34 from 76.73% to 84.16% and DenseNet-121 from 76.73% to 84.16%. The PMD filter yielded limited improvement and slightly reduced the MobileNet-V2 performance. Novelty: This research provides a systematic comparison of contrast enhancement and noise reduction techniques across CNN architectures. This research demonstrates that contrast enhancement is more effective than noise reduction in improving CNN performance. The research provides new pipelines for improving cervical cancer classification.

INTRODUCTION

The introduction motivates automated Pap smear classification and identifies image preprocessing as an underexplored factor. It compares CLAHE and PMD across three CNN architectures, reporting stronger accuracy with CLAHE.

INTRODUCTION

  • Manual Pap smear evaluation depends on specialist expertise and is susceptible to diagnostic variability and human error.The introduction frames automated screening as a response to the growing need for large-scale, reliable screening.Source: INTRODUCTION · p. 1
  • The study situates ResNet-34, DenseNet-121, and MobileNet-V2 among CNN approaches used for medical image classification.ResNet-34 uses residual learning, DenseNet-121 connects layers to facilitate feature reuse, and MobileNet-V2 is described as lightweight.Source: INTRODUCTION · p. 1
  • Preprocessing remains inadequately explored in cervical Pap smear classification, where prior work has emphasized model architectures and related strategies.Comparisons across multiple CNN architectures under identical preprocessing conditions are scarce, and the relative benefits of contrast enhancement and noise reduction remain unclear.Source: INTRODUCTION · p. 2
  • Pap smear images vary in staining, lighting, and preparation, producing low contrast and noise that can obscure cell features important for classification.The study evaluates PMD for noise reduction and edge preservation and CLAHE for local contrast enhancement.Source: INTRODUCTION · p. 2
  • CLAHE outperformed PMD and original images across three CNN architectures, with highest accuracies of 84.16% for ResNet-34, 83.17% for DenseNet-121, and 84.65% for MobileNet-V2.PMD provided only moderate improvements and slightly reduced MobileNet-V2 performance.Source: INTRODUCTION · p. 2

METHODS Proposed methods

The proposed workflow preprocesses SIPaKMeD Pap smear images with CLAHE or PMD before training CNNs to classify cervical cell types. Performance is assessed with a confusion matrix and compared across preprocessing methods.

METHODS Proposed methods

  • The workflow applies either CLAHE for local contrast enhancement or PMD for noise reduction and edge preservation before CNN classification.ResNet-34, MobileNet-V2, and DenseNet-121 classify cervical cell types, with performance assessed using a confusion matrix and compared across preprocessing methods.Source: METHODS Proposed methods · p. 3
  • SIPaKMeD supplies 4,049 cropped single-cell images from 966 original multi-cell microscopic images, annotated into five cell classes.Images were captured at 400× magnification; the dataset does not represent all possible cervical cell variations.Source: METHODS Proposed methods · p. 3
  • The dataset contains five cytopathologist-annotated classes: SI, PB, K, D, and M.These represent superficial-intermediary, parabasal, koilocytotic, dyskeratotic, and metaplastic cells, respectively.Source: METHODS Proposed methods · p. 3
  • Each dataset image is RGB at 256 × 256 pixels, with approximately equal sample counts per class.The images vary in staining intensity, cellular overlap, and background noise.Source: METHODS Proposed methods · p. 3

Contrast enhancement

Image contrast enhancement addresses faint distinctions in Pap smear morphology, while CLAHE improves local contrast and the PMD filter reduces noise while preserving edges.

Contrast enhancement

  • Insufficient contrast can obscure essential cellular morphology, so enhancing contrast can reveal subtle details needed for classification.Contrast describes the visual difference between objects such as a cell’s nucleus and cytoplasm.Source: Contrast enhancement · pp. 3–4
  • CLAHE applies histogram equalization to small image tiles and blends their results to improve local contrast without tile-boundary discontinuities.Unlike standard histogram equalization, CLAHE clips histogram bins to limit noise amplification.Source: Contrast enhancement · p. 4
  • The clip limit caps histogram-bin heights, and clipped surplus is redistributed across the bins.This contrast-limiting step is intended to prevent noise amplification.Source: Contrast enhancement · p. 4
  • Bilinear interpolation between adjacent tiles avoids discontinuities at tile borders.Source: Contrast enhancement · p. 4
  • The PMD filter selectively smooths homogeneous regions while preserving edges, reducing noise without blurring important cellular boundaries.Its diffusion coefficient depends on gradient magnitude, and the contrast-threshold parameter K controls edge sensitivity.Source: Contrast enhancement · pp. 4–5

CNN architectures

The study compares ResNet-34, MobileNet-V2, and DenseNet-121 to assess preprocessing across different network designs and computational complexities.

CNN architectures

  • The study uses ResNet-34, MobileNet-V2, and DenseNet-121 to evaluate preprocessing effects across distinct network designs and computational complexities.Source: CNN architectures · p. 5
  • ResNet-34 uses residual learning, DenseNet-121 uses dense feature connectivity for feature reuse, and MobileNet-V2 uses depth-wise separable convolutions as a lightweight design.Source: CNN architectures · p. 5
  • All three models began with ImageNet-pretrained weights and were trained using identical experimental settings.Source: CNN architectures · p. 5

Simulation setting

The preprocessing pipeline applies CLAHE or PMD to each RGB channel before recombining the channels, and compares models under shared training conditions.

Simulation setting

  • Figure 2 depicts the image preprocessing step using either the PMD filter or CLAHE.Source: Simulation setting · p. 5, Figure 2 · p. 6
  • CLAHE was configured with block_size = 8 × 8 and clip_limit = 2.0, while PMD used N = 20, λ = 0.25, and κ = 20.Each method was applied independently to the RGB channels, which were recombined after enhancement.Source: Simulation setting · pp. 5–6
  • Models were trained for 30 epochs with Adamax, a 0.001 learning rate, batch size 32, and an 80%/10%/10% train/test/validation split.All models used identical experimental conditions; a fixed random seed of 42 was used throughout.Source: Simulation setting · p. 6

Performance metrics

Model performance was assessed with a confusion matrix and four metrics calculated for each class and then averaged.

Performance metrics

  • A confusion matrix was used to assess model performance by comparing predicted classes with actual classes.Source: Performance metrics · p. 7
  • Accuracy, recall, precision, and F1-score were derived from the confusion-matrix values.Source: Performance metrics · p. 7
  • The metrics were calculated for each class in the multiclass setting and then averaged to summarize overall performance and class-specific difficulties.The confusion matrix can also identify frequently confused cell types.Source: Performance metrics · p. 7

RESULT AND DISCUSSION

The visual comparison shows that CLAHE improves local contrast and cellular detail, whereas PMD reduces noise but softens some fine textures and edges.

Qualitative visual inspection

  • Figure 3 presents original images in column (a), CLAHE-enhanced images in (b), and PMD-filtered images in (c).Source: RESULT AND DISCUSSION · p. 7, Figure 3 · p. 7
  • CLAHE makes nuclear boundaries sharper and cellular detail easier to interpret than in the original images.The original images have low contrast and subtle noise that obscure cellular structures, especially at their boundaries.Source: RESULT AND DISCUSSION · p. 7, Figure 3 · p. 7
  • PMD smooths images and reduces noise while preserving general cell shape, but can soften fine textures and edges.The passage describes PMD as a complementary noise-reduction technique that does not improve contrast as effectively as CLAHE.Source: RESULT AND DISCUSSION · p. 7

ResNet-34 performances

For ResNet-34, preprocessing improved performance, with CLAHE producing larger gains across metrics than PMD. The reported PMD metrics are above the original-image baseline.

ResNet-34 performances

  • 76.73% accuracy, 76.90% recall, 76.19% F1-score, and 79.12% precision were the original-image baseline metrics.Source: ResNet-34 performances · p. 8
  • CLAHE substantially improved all ResNet-34 performance metrics, while PMD gains were less substantial than CLAHE's.Figure 4 shows the effect of preprocessing on ResNet-34 performance.Source: ResNet-34 performances · p. 8, Figure 4 · p. 8
  • 79.21% accuracy, 80.49% precision, 79.39% recall, and 78.76% F1-score were attained with PMD-preprocessed images.The passage reports that PMD improved over the original dataset, though less than CLAHE.Source: ResNet-34 performances · p. 8

MobileNet-V2 performance

For MobileNet-V2, CLAHE yielded modest improvement, whereas PMD slightly degraded performance. The original-image baseline was already strong, and the reported PMD accuracy was lower than that baseline.

MobileNet-V2 performance

  • 83.91% accuracy was achieved by MobileNet-V2 on original images, which the passage describes as a strong baseline.Source: MobileNet-V2 performance · p. 8
  • 84.65% accuracy and 84.17% F1-score were reported for MobileNet-V2 with CLAHE preprocessing.Figure 5 presents the effect of image preprocessing on MobileNet-V2 performance.Source: Figure 5 · p. 8
  • 81.93% accuracy followed PMD filtering, a slight degradation associated with suppression of subtle high-frequency features important for classification.The passage suggests the feature loss may affect MobileNet-V2 more noticeably because of its depthwise separable convolutions and lighter parameterization.Source: MobileNet-V2 performance · p. 9
  • The reported findings characterize CLAHE as more effective than noise reduction for optimizing MobileNet-V2 performance.The passage also cautions that PMD may compromise the feature richness needed by lightweight architectures.Source: MobileNet-V2 performance · p. 9

DenseNet-121 performances

DenseNet-121 performed best with CLAHE, while PMD produced moderate gains over original images. The passages attribute CLAHE’s advantage to increased local contrast and clearer cellular morphology.

DenseNet-121 performances

  • CLAHE increases local contrast, making cervical-cell morphology—including size, texture, and boundary details—more distinguishable to DenseNet-121.The discussion also identifies nuclear shape and nuclear-cytoplasmic ratio as subtle features relevant to Pap smear classification.Source: DenseNet-121 performances · pp. 9–10
  • 83.17% accuracy with CLAHE, versus 76.73% on original images and 78.22% with PMD, made CLAHE the strongest DenseNet-121 preprocessing option.On original images, precision was 79.81%, recall 76.85%, and F1-score 76.04%; with CLAHE, these were 84.85%, 83.40%, and 82.56%, respectively.Source: Figure 6 · p. 9, DenseNet-121 performances · p. 9
  • PMD achieved 78.22% accuracy, 82.73% precision, 78.47% recall, and 77.42% F1-score, improving over original images but trailing CLAHE overall.The discussion says PMD reduces noise while preserving major boundaries, but does not substantially increase contrast and may attenuate fine textures and high-frequency details.Source: DenseNet-121 performances · pp. 9–10
  • Preprocessing effects varied by architecture; DenseNet-121 benefits from CLAHE, with dense connectivity supporting feature reuse and information flow.The discussion contrasts this with MobileNet-V2’s slight decline after PMD preprocessing.Source: DenseNet-121 performances · p. 10

CONCLUSION

The study compares CLAHE and PMD for CNN-based cervical cancer classification on SIPaKMeD. CLAHE achieved higher accuracy than original images across all three evaluated architectures, while the study is limited to one dataset and two preprocessing methods.

CONCLUSION

  • CLAHE increased accuracy from 76.73% to 84.16% for ResNet-34, 76.73% to 83.17% for DenseNet-121, and 83.91% to 84.65% for MobileNet-V2.Source: CONCLUSION · p. 10
  • PMD provided moderate improvements overall but slightly reduced MobileNet-V2 performance.Source: CONCLUSION · p. 10
  • The comparison covers two preprocessing methods on one dataset; future work will examine optimized CLAHE, hybrid pipelines, and additional cervical cytology datasets.Source: CONCLUSION · p. 10
Loading 2610.11086v1…