PaperlayerView exact revision on arXiv

Source-linked AI summary

Comprehensive Benchmarking of Deep Learning Architectures for Lung Cancer Histopathology

Hadi Hasan, Safaa Salman, Lama Sleem, Ralph Mouawad, Ali Chehab

arXiv:2608.15915v1cs.CVcs.AI

TL;DR

Manual lung histopathology is labor-intensive and subject to inter-observer variability, while integrated benchmarks spanning classification and segmentation remain limited. This study benchmarks and integrates deep learning architectures across both tasks, finding YOLO11 best for classification and DeepLabV3+ best for segmentation, with YOLO11-seg offering comparable efficiency.

  • Problem

    Integrated evidence comparing tissue classification and spatial region segmentation for lung histopathology remains limited despite manual diagnosis being labor-intensive and variable.

  • Method

    The study systematically benchmarks six classifiers and four segmentation models under a controlled protocol, then cascades the best models into a two-stage analysis pipeline.

  • Results

    YOLO11 leads classification at 98.38% accuracy, while DeepLabV3+ leads segmentation at IoU 0.80 and YOLO11-seg achieves IoU 0.79 with 14× fewer parameters.

  • Takeaways & Limitations

    The benchmark provides a comparative baseline showing strong classification performance and a segmentation accuracy–efficiency trade-off for automated lung histopathology analysis.

Abstract

from arXiv · show

Lung cancer remains the leading cause of cancer-related mortality worldwide, while histopathological diagnosis is often affected by inter-observer variability and the substantial workload associated with manual slide examination. Although deep learning has shown considerable potential in computational pathology, comprehensive benchmarks that integrate tissue classification and region segmentation within a unified analytical framework remain limited. This study presents a two-stage deep learning framework for multi-class tissue classification and pixel-level histopathological region segmentation, accompanied by a systematic comparison of state-of-the-art architectures at each stage. For tissue classification, six models, a custom convolutional neural network, VGG16, DenseNet, MobileNetV3, a custom Vision Transformer, and YOLO11, are evaluated on a combined dataset of 39,000 images derived from LC25000 and LungHist700. The models distinguish between adenocarcinoma, squamous cell carcinoma, and normal lung tissue. YOLO11 achieves the best classification performance, with an accuracy of 98.38%, a five-fold cross-validation accuracy of 98.21 +/- 0.35%, and a macro F1-score of 0.98. For region segmentation, U-Net, ResNet-encoder U-Net, DeepLabV3+, and YOLO11-seg are evaluated using the GlaS gland segmentation benchmark. DeepLabV3+ obtains the highest Intersection over Union of 0.80 and a Dice score of 0.89, while YOLO11-seg achieves a comparable Intersection over Union of 0.79 using approximately 14x fewer parameters. The best-performing classification and segmentation models are subsequently integrated into an end-to-end framework, providing an accurate, computationally efficient, and reproducible baseline for automated histopathological image analysis.

I. INTRODUCTION

Lung histopathology diagnosis is clinically important but labor-intensive and variable, motivating integrated automation of tissue classification and pathological-region delineation. This study addresses benchmarking gaps through controlled comparisons of established architectures across both tasks, emphasizing reproducibility and accuracy–efficiency trade-offs.

  • Clinical motivation: Histopathology is the diagnostic gold standard, but lung-cancer analysis depends on expertise and is affected by inter-observer variability, cognitive burden, and labor-intensive slide examination.Variability also arises from staining protocols, tissue preparation, and intra-tumoral heterogeneity.
  • Clinical motivation: Automated lung histopathology must address both tissue classification—ACA, SCC, or normal—and spatial delineation of pathological regions within an integrated framework.The passages identify integrated automation of both tasks as having considerable clinical value.
  • Research gaps: Deep learning offers consistent, reproducible predictions at scale, yet most studies treat classification and segmentation separately and rarely compare convolutional, transformer-based, and detection-based architectures under one protocol.These gaps complicate principled model selection.
  • Study contribution: The study benchmarks six classifiers and four segmentation models under fixed preprocessing, augmentation, optimization, data partitioning, cross-validation, and hardware conditions.The controlled design attributes observed differences to architecture rather than training-recipe variation.
  • Study contribution: The benchmark reports per-class metrics, stratified 5-fold cross-validation, parameter count, GFLOPs, and measured latency for every configuration.Model selection is framed as an accuracy–efficiency trade-off rather than a leaderboard ranking.

II. CLINICAL BACKGROUND · III. RELATED WORK

Lung histopathology distinguishes clinically consequential NSCLC subtypes through microscopic or digitized tissue examination. Prior work spans CNN classification, transformer and foundation-model representations, segmentation architectures, and relatively scarce integrated classification–segmentation pipelines.

  • II. CLINICAL BACKGROUND: Histopathology examines biopsy or resection tissue after fixation, paraffin embedding, thin sectioning, and usually hematoxylin-and-eosin staining.H&E renders nuclei blue-purple and cytoplasm and stroma pink before microscopic or whole-slide-image examination.
  • II. CLINICAL BACKGROUND: Adenocarcinoma typically shows gland formation or lepidic growth, whereas squamous cell carcinoma is recognized by keratinization and intercellular bridges.The two are dominant NSCLC subtypes with distinct epithelial origins.
  • II. CLINICAL BACKGROUND: Subtype assignment directly conditions eligibility for targeted therapy, immunotherapy, and specific chemotherapeutic agents.This makes distinguishing adenocarcinoma from squamous cell carcinoma consequential for treatment rather than merely taxonomic.
  • III. RELATED WORK: 97.3–99.9% F1-scores were reported across six pretrained CNNs on the three-class LC25000 lung subset, while VGG16 reached 98.83% through transfer learning.Earlier approaches also included task-specific CNNs and CNN–handcrafted-feature fusion.
  • III. RELATED WORK: Vision Transformers model long-range whole-slide spatial dependencies, while UNI, CONCH, Virchow, and Prov-GigaPath extend representation learning through large-scale or specialized pretraining.The cited models include tissue-patch, image–text, pan-cancer, and whole-slide-context approaches.
  • III. RELATED WORK: Large pathology foundation models and promptable segmentation models define a representational upper bound but require pretraining corpora and inference budgets unavailable to most laboratories.The benchmark therefore focuses on publicly available, ImageNet-scale architectures.
  • III. RELATED WORK: U-Net established the dominant biomedical segmentation paradigm through skip connections, while deeper pretrained encoders such as ResNet substantially improve segmentation quality.Fully convolutional networks established the encoder–decoder foundation; a U-Net-inspired gland model reported a Jaccard index of 86.4%.
  • III. RELATED WORK: Integrated lung-histopathology pipelines cascading tissue classification with region segmentation remain scarce, although decoupled multi-stage structures are established in other clinical-imaging applications.Earlier work used cascading detection and downstream characterization for opportunistic vertebral-fracture and Schmorl’s-node screening from routine CT.

IV. METHODOLOGY

The methodology covers datasets, preprocessing, model architectures, and training strategies within a cascaded pipeline. The pipeline combines tissue classification and segmentation for end-to-end histopathology slide analysis.

  • The methodology includes datasets, preprocessing, model architectures, and training strategies.
  • The overall pipeline cascades a tissue classifier with a segmentation model.
  • The cascade progresses from tissue-type prediction to spatial delineation of affected regions, enabling end-to-end histopathology slide analysis.

A. Datasets

The study uses LC25000, LungHist700, and GlaS for class-level classification and pixel-level segmentation. LungHist700 is expanded through non-overlapping cropping and merged with LC25000 into a 39,000-image, three-class classification corpus, while GlaS supplies segmentation annotations despite its colorectal domain.

  • Dataset composition: 25,000 LC25000 images, 700 LungHist700 images, and 165 GlaS images form the study’s three public histopathology datasets.LC25000 images are 768 × 768, LungHist700 images are 1600 × 1200, and GlaS images are 775 × 522.
  • Dataset composition: 14,000 non-overlapping patches are generated from the 700 LungHist700 images to approximate LC25000’s field of view.Parent-image-level stratified splitting prevents leakage between training and test sets because adjacent crops share morphological context.
  • Classification corpus: 39,000 images span ACA, SCC, and NOR after merging LC25000 with the expanded LungHist700 classification corpus.The first two datasets provide class-level labels.
  • Segmentation dataset: 165 pixel-level annotated GlaS images support supervised gland segmentation, although the colorectal benchmark is not lung-specific.The dataset choice reflects the scarcity of lung-specific pixel-level annotations and is identified as a study limitation.

B. Data Preprocessing · C. Classification Models · D. Segmentation Models

The framework standardizes histopathology images through resizing, stain normalization, augmentation, and stratified splitting, then benchmarks six classification and four segmentation architectures under pretrained or task-specific configurations. YOLO11 is adapted for classification, while DeepLabV3+ uses multi-scale contextual encoding and fine-detail decoding for tissue-region segmentation.

  • B. Data Preprocessing: 768 × 768 pixels standardization and per-channel z-normalization reconciled image dimensions and staining differences between LC25000 and LungHist700.Normalization used channel-wise mean subtraction and standard-deviation scaling.
  • B. Data Preprocessing: Random rotation (±90◦), horizontal and vertical flipping, brightness and contrast jittering (±20%), and elastic deformation expanded training diversity.These augmentations were applied to reduce overfitting.
  • B. Data Preprocessing: An 80/10/10 training-validation-test split used stratified parent-image sampling to preserve class balance, with pixel values normalized to [0, 1].The split followed augmentation and pixel normalization preprocessing steps.
  • C. Classification Models: YOLO11 replaced its detection head with a fully connected classification layer and fine-tuned an ImageNet-pretrained convolutional backbone on the merged dataset.Its C3k2, C2PSA, and spatial pyramid pooling modules aggregate features across multiple spatial scales.
  • C. Classification Models: Six classification architectures were evaluated: a custom CNN, VGG16, DenseNet, MobileNetV3, a custom Vision Transformer, and YOLO11.The supplied passage explicitly describes YOLO11 and lists the additional architectures benchmarked alongside it.
  • D. Segmentation Models: DeepLabV3+ combined an atrous-convolution encoder, ASPP module, and lightweight decoder to capture multi-scale context and recover fine spatial detail.An ImageNet-pretrained ResNet-101 encoded GlaS images, while the ASPP module and decoder were fine-tuned on the dataset.
  • D. Segmentation Models: Four segmentation architectures were trained under the same conditions: U-Net, ResNet-encoder U-Net, DeepLabV3+, and YOLO11 instance-segmentation.The architectures were evaluated for irregular, heterogeneous tissue-region segmentation on GlaS.

E. Training Strategy … B. Classification Results

The study used transfer learning and standardized optimization across most architectures, then compared six classifiers on held-out and cross-validated performance. YOLO11 achieved the strongest classification results, while the from-scratch custom CNN underfit substantially.

  • E. Training Strategy: ImageNet-pretrained models were fine-tuned with transfer learning, whereas YOLO11-seg used COCO segmentation weights.The custom CNN used a lower initial learning rate of 0.0001 to stabilize training of its randomly initialized weights.
  • E. Training Strategy: Adam with cosine annealing and an initial learning rate of 0.001 was used for all models except the custom CNN.The custom CNN instead used an initial learning rate of 0.0001.
  • A. Experimental Setup: Experiments ran on an NVIDIA Tesla P100 GPU with 32 GB RAM using Python, TensorFlow, and PyTorch.Primary results used an 80/10/10 held-out split, with stratified 5-fold cross-validation reported alongside them.
  • B. Classification Results: 98.38% was YOLO11’s best overall classification accuracy, while the custom CNN capped at 83.5%.The six classifiers were evaluated using single-split accuracy, macro F1-score, and 5-fold cross-validation means.
  • V. EXPERIMENTS AND RESULTS: The standardized configuration used 768 × 768 RGB inputs, early stopping at patience 3, a maximum of 100 epochs, and stratified 5-fold cross-validation.The table also records each model’s initialization status, weight decay, and cross-entropy settings.
  • B. Classification Results: ±0.35% was YOLO11’s smallest cross-validation standard deviation, indicating stable performance across data partitions.Its cross-validated means closely aligned with the single-split results.
  • B. Classification Results: 97.1% was achieved by the custom Vision Transformer despite training from scratch, while DenseNet and MobileNetV3 performed almost identically.MobileNetV3 delivered this comparable classification performance at much lower computational cost, and VGG16 trailed denser networks.

1) Per-Class Analysis:

Across all six models, normal lung tissue is easiest to distinguish, whereas ACA–SCC classification is consistently hardest because of morphological overlap. YOLO11 provides the strongest and most balanced per-class performance, especially for SCC.

  • Per-Class Analysis:: Normal tissue is the most readily distinguished class, while the clinically critical ACA vs. SCC distinction is consistently hardest because of morphological overlap.
  • Per-Class Analysis:: F1 = 0.80 for CNN and 0.97 for YOLO11 define the SCC range across models, with SCC the weakest class for every model.
  • Per-Class Analysis:: ACA 0.98, SCC 0.97, and NOR 0.99 give YOLO11 balanced performance across all three classes.Its lowest per-class F1 matches or exceeds the macro F1 of every other evaluated model.

C. Segmentation Results

DeepLabV3+ achieved the strongest segmentation performance on the GlaS test set, while encoder pretraining substantially improved U-Net results. Qualitative outputs similarly showed fragmented vanilla U-Net masks versus coherent boundaries from DeepLabV3+ and YOLO11-seg.

  • Quantitative performance: IoU = 0.80 and Dice = 0.89 made DeepLabV3+ the best-performing segmentation architecture.Its ASPP-based multi-scale aggregation supported segmentation across glands of varying sizes.
  • Encoder effects: 0.27 IoU was gained by ResNet-U-Net over vanilla U-Net through pretrained encoder substitution.Vanilla U-Net used a limited non-pretrained encoder trained on only 165 images.
  • Quantitative performance: ±0.02 IoU was DeepLabV3+’s lowest cross-validation variance.The reported results used five-fold cross-validation on the GlaS test set.
  • Qualitative assessment: DeepLabV3+ and YOLO11-seg recovered coherent gland boundaries despite very different parameter budgets.DeepLabV3+ yielded the cleanest boundaries, while YOLO11-seg occasionally merged regions because it predicted per-instance polygon masks rather than a dense pixel map.
  • Qualitative assessment: Vanilla U-Net produced fragmented, jagged, discontinuous masks, whereas pretrained ResNet-U-Net restored coherent glandular contours.The qualitative differences were consistent with the quantitative ranking.

D. Computational Cost

At 768 × 768 resolution, the benchmark compares model parameters, GFLOPs, and single-image inference latency. YOLO11 provides the strongest classifier accuracy–efficiency trade-off, while DeepLabV3+ achieves the best segmentation IoU at the highest cost.

  • Classification cost: 9.4M parameters, 13.5 GFLOPs, and 6 ms latency make YOLO11 the most efficient classifier with the best accuracy–efficiency trade-off.Measurements are reported for single-image inference at 768 × 768 resolution.
  • Classification cost: 134.3M parameters make VGG16 by far the most parameter-heavy model, yet it delivers only moderate accuracy.MobileNetV3 reaches higher accuracy with 25× fewer parameters than VGG16.
  • Segmentation cost: DeepLabV3+ incurs the highest segmentation cost but yields the best IoU.YOLO11-seg attains nearly equivalent IoU at lower computational cost.

E. Ablation Studies · F. Comparison with Prior Work · VI. DISCUSSION

Ablation studies identify merged data, augmentation, and encoder choice as important design factors, while prior-work comparisons emphasize differing benchmark setups. Discussion findings highlight transfer learning, cross-domain cascade transfer, and clinical-use limitations involving adversarial robustness and uncertainty.

  • E. Ablation Studies: 98.38% accuracy on the merged corpus exceeds LC25000-only training by 0.56 percentage points and LungHist700-only training by 6.95 points.The respective single-source accuracies are 97.82% and 91.43%.
  • E. Ablation Studies: 1.47 points: disabling augmentation reduces YOLO11 accuracy from 98.38% to 96.91%, demonstrating its contribution to robustness.
  • E. Ablation Studies: 0.80 IoU and 0.89 Dice: DeepLabV3+ with a ResNet-101 encoder outperforms MobileNetV2 at 0.74 IoU and ResNet-50 at 0.78 IoU on GlaS.The passage describes diminishing returns from deeper encoders on this small dataset.
  • F. Comparison with Prior Work: 98.38%: YOLO11’s merged three-class result differs from prior studies using LC25000 alone or a full five-class version, limiting direct comparability.Reported prior results include 97.3–99.9% F1-scores, 98.83% with VGG16, and 99.39% on the full five-class version.
  • VI. DISCUSSION: ∼15 percentage points: the gap between the non-pretrained custom CNN and YOLO11 underscores transfer learning’s role, while progression through VGG16, DenseNet, and MobileNetV3 compounds that benefit.
  • VI. DISCUSSION: The decoupled two-stage cascade transfers across domains because its stages share preprocessing and evaluation protocols, not dataset-specific tuning.The structure is instantiated for routine CT screening of opportunistic vertebral fracture and Schmorl’s node.
  • VI. DISCUSSION: Absolute figures should not be assumed to carry over across modalities, and another modality would require re-implementation of the approach.
  • VI. DISCUSSION: Clinical use is conditioned by vulnerabilities that conventional metrics do not surface, including adaptive attacks, while reinforcement-learning-based hardening is orthogonal to reported accuracy, IoU, and latency.

VII. CONCLUSION

The study contributes a systematic multi-architecture benchmark and integrated two-stage pipeline for automated lung cancer histopathology classification and segmentation. Under a single controlled protocol, YOLO11 achieved the strongest reported classification performance on the merged dataset.

  • VII. CONCLUSION: The work presents a systematic multi-architecture benchmark and integrated two-stage pipeline for automated lung cancer histopathology classification and segmentation.Its contribution is comparative evidence gathered under a single controlled protocol rather than a new architecture.
  • VII. CONCLUSION: 98.38% test accuracy: YOLO11 performed best among six classifiers on the merged 39,000-image LC25000–LungHist700 dataset.The benchmark compares six classifiers under the same evaluation protocol.
  • VII. CONCLUSION: 98.21 ± 0.35% 5-fold CV accuracy: YOLO11 maintained the leading classification result across cross-validation.The reported cross-validation result accompanies the 98.38% test accuracy.
Loading 2608.15915v1…