Lung cancer causes approximately 1.76 million fatalities and 2.06 million new cases annually worldwide, while colorectal cancer accounts for 783,000 deaths and 1.80 million new cases. Accurate and timely histopathological diagnosis is essential for selecting appropriate treatment, as therapeutic choices depend on cancer subtype, stage, and molecular characteristics. The global shortage of skilled histopathologists creates a critical bottleneck in diagnostic throughput and consistency.
Conventional histopathological diagnosis requires tedious manual evaluation by expert pathologists, which introduces subjectivity, inter-observer variability, and diagnostic delays. A single pathologist may examine hundreds of image slides, increasing the risk of cognitive fatigue-related errors. Automating the identification of cancerous tissue patterns through deep learning offers a pathway to standardized, rapid, and objective cancer classification without dependence on scarce expert resources.
Low image resolution is a fundamental technical obstacle in histopathological analysis, limiting the visibility of subtle cellular morphological features that distinguish cancer subtypes. Standard digitization of tissue slides can produce images with insufficient detail to clearly differentiate adenocarcinoma cells from squamous cell carcinoma or benign tissue. Super-resolution preprocessing that enhances image clarity before classification represents a novel approach to improving diagnostic accuracy.
This study introduces a two-stage framework combining Fast Super-Resolution CNN (FSRCNN) for image enhancement with DenseNet201 deep feature extraction and SVM classification on the LC25000 histopathological image dataset. The approach addresses both the image quality limitation and the feature discriminability limitation simultaneously, testing whether this dual enhancement strategy yields meaningful improvements over existing CNN-only approaches.
The LC25000 dataset contains 25,000 color histopathological images at 768x768 pixel resolution, organized into five classes of 5,000 images each: colon adenocarcinoma, benign colon tissue, lung adenocarcinoma, benign lung tissue, and lung squamous cell carcinoma. The dataset was split 70% for training (3,501 images per class), 20% for testing (1,000 images per class), and 10% for validation (500 images per class). Images were normalized to the 0-1 range using per-dataset mean and standard deviation.
FSRCNN is an improved super-resolution network that operates on low-resolution input images directly, avoiding the bicubic upsampling preprocessing required by earlier SRCNN architectures. It consists of five stages: feature extraction via 5x5 convolution, shrinking via 1x1 convolution to reduce feature map dimensionality, nonlinear mapping via multiple 3x3 convolutional layers, expanding back to full feature map count, and deconvolution using 9x9 filters to reconstruct the high-resolution output. FSRCNN achieves similar restoration quality to SRCNN at up to 40 times faster inference speed due to its lower-resolution processing pipeline.
DenseNet201, a 201-layer dense convolutional network derived from ResNet, connects each layer directly to all preceding layers within dense blocks, promoting feature reuse and efficient gradient flow throughout training. The architecture uses three transition layers and four dense blocks with 1x1 and 3x3 convolutional kernels. A dropout rate of 0.9 was applied to prevent overfitting, substantially outperforming lower dropout rates of 0.1 through 0.7 which were also tested. Input images were resized to 224x224x3 pixels to match DenseNet201 requirements.
Deep features extracted from the fc1000 fully connected layer of DenseNet201 were passed to a Support Vector Machine classifier that created a decision boundary in the high-dimensional feature space for final cancer type discrimination. SVM was selected for its established effectiveness in high-dimensional data scenarios and its margin-based classification approach that tends to generalize well to unseen data. Training used the Adam optimizer with a learning rate of 0.0001 over 30 epochs with a batch size of 4, balancing computational efficiency and memory requirements.
The proposed FSRCNN-DenseNet201-SVM model achieved an overall accuracy of 98.00%, precision of 98.10%, sensitivity of 98%, specificity of 99.50%, and F1 score of 0.98 across all five tissue categories. These results were achieved with a 0.9 dropout rate, which consistently outperformed all other tested dropout rates. The high specificity of 99.50% is particularly notable for clinical applications where false positives lead to unnecessary treatments and patient anxiety.
Class-specific performance varied across the five tissue categories, with benign tissue classes showing the highest individual scores. Benign colon tissue achieved an F1 score of 0.999 with zero false positives and 99.80% sensitivity. Benign lung tissue achieved an F1 score of 0.999 with near-zero false positive rate and 100% sensitivity. Lung squamous cell carcinoma achieved an F1 score of 0.9975 with 100% precision and 99.50% sensitivity, indicating near-perfect discrimination of this subtype.
The cancer subtypes proved more challenging to classify, with lung adenocarcinoma achieving an F1 score of 0.9533 and colon adenocarcinoma achieving 0.9513, both with higher false positive rates than benign classes. Lung adenocarcinoma showed 92.00% precision but 98.90% sensitivity, indicating the model was slightly more likely to over-call lung adenocarcinoma than to miss it. This pattern is clinically acceptable since missed cancer is generally more harmful than additional confirmatory testing triggered by a false positive.
The model contains approximately 20 million trainable parameters and achieves an average inference time of 100 milliseconds per image, making it practical for clinical settings where real-time or near-real-time results are required. This computational profile is achievable on standard hardware including the NVIDIA GTX3050 used in this study, suggesting that deployment would not require specialized or expensive computing infrastructure beyond what many medical centers already possess.
FSRCNN addresses a fundamental limitation in digitized histopathology: the inevitable resolution loss during slide scanning and digital transmission that can obscure the subtle cellular features most relevant to cancer subtype discrimination. By reconstructing high-resolution image details from lower-resolution inputs, FSRCNN restores the fine-grained texture and nuclear morphology information that distinguishes adenocarcinoma glandular patterns from squamous cell carcinoma keratinization patterns. This preprocessing step ensures that DenseNet201 receives maximally informative input.
DenseNet201's architecture advantage over traditional CNNs lies in its dense connectivity pattern where each layer receives feature maps from all preceding layers as concatenated inputs rather than just the immediately previous layer. This design promotes feature reuse across the network depth, reduces the number of parameters needed compared to equivalent traditional architectures, and facilitates gradient flow during training that prevents the vanishing gradient problem common in very deep networks. The result is a model that learns both low-level texture features and high-level structural patterns relevant to cancer classification.
The combination of DenseNet201 feature extraction with SVM classification rather than a fully connected softmax classifier harnesses complementary algorithmic strengths. DenseNet201 excels at learning hierarchical image representations but its built-in softmax output is less optimal than SVM for problems requiring strong margin separation in high-dimensional feature spaces. SVM's margin maximization principle encourages the classifier to find the most generalizable boundary between classes, reducing overfitting risk on the highly structured features extracted by DenseNet201.
Dropout regularization at 0.9, meaning 90% of neurons are deactivated during each training pass, provides aggressive regularization that prevents the model from memorizing training images. The systematic comparison of five dropout rates (0.1, 0.3, 0.5, 0.7, 0.9) showed monotonically improving accuracy as dropout increased from 95.0% to 98.0%, confirming that the LC25000 dataset is large enough to benefit from strong regularization rather than being harmed by reduced capacity during training.
Compared to prior methods on the LC25000 dataset, the proposed model achieves competitive performance while providing superior specificity and a more complete set of clinical evaluation metrics. The highest-accuracy prior approach (EfficientNetV2 at 99.96%) did not report specificity or sensitivity, making direct clinical comparability impossible. A model optimized purely for accuracy can achieve high scores by correctly classifying the dominant class while performing poorly on minority classes, an issue that balanced reporting with specificity and sensitivity helps reveal.
The inclusion of FSRCNN provides a key advantage in real-world deployment scenarios where image quality varies across scanning equipment, digitization protocols, and storage formats. Models trained and tested on clean, high-resolution images may degrade significantly when deployed on images from different scanners or compressed during storage and transmission. FSRCNN preprocessing creates a consistent high-resolution representation regardless of input quality, potentially improving robustness across diverse clinical environments.
The hybrid DenseNet201-SVM architecture outperforms pure CNN approaches by combining the representation learning strength of deep networks with the classification robustness of kernel-based methods. This hybrid strategy is particularly relevant for medical imaging applications where training datasets, while large by histopathology standards at 25,000 images, are still small compared to natural image datasets. SVM's effective inductive bias for high-dimensional low-sample problems complements DenseNet201's data-hungry feature learning.
The study acknowledges that LC25000, while providing a solid foundation, represents a curated research dataset that may not capture the full variability of real clinical histopathology images including staining artifacts, tissue processing variations, and rare morphological presentations. Validation on diverse multi-institution datasets with variable staining protocols and scanning equipment is necessary before clinical deployment. Integration of genomic markers and electronic health record data in future multimodal approaches could further improve diagnostic precision beyond what image analysis alone can achieve.
This study demonstrates that combining FSRCNN super-resolution image enhancement with DenseNet201 deep feature extraction and SVM classification achieves 98.00% accuracy, 98.10% precision, 98% sensitivity, and 99.50% specificity for distinguishing five histopathological tissue categories covering lung and colon cancers. The two-stage approach addresses both the image quality limitation and the feature discriminability challenge that have constrained prior single-stage approaches, resulting in a robust and clinically relevant diagnostic framework.
The consistent performance across all five tissue categories, including the two more challenging cancer subtypes, confirms that the model has learned generalizable discriminative features rather than class-specific memorization. The high specificity of 99.50% is particularly important for clinical applications, as minimizing false positives is essential for preventing unnecessary treatment, reducing patient anxiety, and maintaining appropriate resource allocation in pathology departments.
Future research should prioritize validation on larger and more diverse clinical datasets from multiple institutions, scanner types, and staining protocols to assess real-world generalizability. Multimodal extensions combining histopathological deep features with genomic markers, patient clinical histories, and imaging from CT or MRI could create more comprehensive diagnostic tools that leverage all available clinical information rather than image data alone.
The practical deployment requirements of the model, including 100ms inference time and compatibility with standard GPU hardware, position this framework as a clinically feasible tool for high-throughput pathology laboratories. Deployment would require integration with existing laboratory information management systems and regulatory approval processes, but the computational requirements do not present a barrier that would prevent implementation in well-resourced clinical environments.