Convolution Neural Networks for diagnosing colon and lung cancer histopathological images

• Cancer is described as a leading cause of mortality, with lung cancer accounting for 11.6% of cases and 18.4% of cancer deaths, and colorectal cancer contributing 9.2% of mortality.Source: §1 · p. 1
• Lung cancer symptoms may not appear until the disease has spread, making treatment harder.Source: §1 · p. 1
• Adenocarcinoma and squamous cell carcinoma are identified as the most common lung cancer types, with small- and large-cell carcinomas also described.Source: §1 · pp. 1–2
• Colon cancer often begins as benign polyps, while colon adenocarcinomas develop in the intestinal lining and spread to other layers.Source: §1 · p. 2
• The study proposes CNNs to detect common lung and colon cancers using histopathology slides, which the authors describe as a gold standard for distinguishing cancer types.The introduction also says CNNs learn image features, reducing reliance on hand-engineered filters and requiring less preprocessing than other classification algorithms.Source: §1 · p. 2
• The paper is organized around prior research, the LC25000 dataset, CNN methods and architecture, experiments, and conclusions.Source: §1 · p. 2

• Earlier computer-aided diagnosis relied on feature extraction, while cancer image analysis has since covered multiple cancer types and deep-learning architectures.Munir et al. reviewed cancer-diagnosis experiments and compared prominent architectures.Source: §2 · p. 2
• Coudray et al. reported a mean AUC of 0.97 using InceptionV3 to classify lung adenocarcinoma and squamous cell carcinoma images.Other cited work predicted lung-cancer risk from CT scans or applied deep networks to CT diagnosis and colon-tissue nuclei classification.Source: §2 · p. 2
• Comparative CNN studies reported an F-1 score of 0.97 for lung squamous cell carcinoma and baseline accuracy of 93% for colonic adenocarcinoma.Source: §2 · p. 2

• Each of LC25000’s five classes—lung adenocarcinoma, lung squamous cell carcinoma, lung benign, colon adenocarcinoma, and colon benign—contains 5,000 images.Source: §3 · p. 3
• The original dataset had 750 lung and 500 colon images at 1024×768 pixels, then images were converted to 768×768 and augmented to 25,000 using rotations and flips.Source: §3 · p. 3
• Figure 1 shows lung adenocarcinoma and squamous cell carcinoma examples, benign lung tissue, colon adenocarcinoma, and benign colon tissue.Source: Figure 1 · p. 3
• For each class, the authors randomly sampled 4,500 training and 500 test images, then resized images to 150×150 and applied randomized shear, zoom, and normalization.Source: §3 · p. 3

• Histopathology classification is challenging because images contain complex, interwoven structures and similar morphological textures.Source: §4 · p. 3
• CNNs learn features directly from input data and stack trainable layers with a supervised classifier to form feature maps.Source: §4 · p. 3
• A typical CNN combines convolution, max-pooling, and fully connected or dense layers; Figure 2 depicts a commonly used architecture.Source: §4 · p. 4, Figure 2 · p. 4
• Convolution filters apply shared weights across local image regions, with stride setting the distance between filter applications.When stride is smaller than the filter size, the convolutions overlap.Source: §4 · p. 4
• Pooling between convolution layers downsamples spatial dimensions, reducing parameters and computation and helping control overfitting.Max pooling selects the maximum value within a neighborhood.Source: §4 · p. 4
• The dense layer connects input activations through matrix multiplication and bias offsets, with the final layer producing probabilities or logits.Source: §4 · p. 4

5 CNN Architecture and Training Strategy

• Images enter the network as 150×150 RGB inputs, then pass through max pooling after each convolution layer.Each pooling layer uses size 2 and valid padding; the flattened output is converted to a 1D tensor before a dense layer.Source: §5 · p. 4
• The CNN uses three convolution layers, with 32 filters in the first and 64 in each of the next two.Each uses 3×3 filters, stride 2, same padding, Gaussian initialization, and ReLU activation.Source: §5 · p. 4
• Two dense layers use 512 neurons in the first layer and 3 lung or 2 colon output neurons in the last.The final dense output can be activated with softmax; dropout between dense layers randomly drops neurons to prevent overfitting.Source: §5 · p. 5
• Table 1 summarizes the CNN layers, with CONV+POOL denoting convolution followed by pooling and FC denoting a fully connected layer.The architecture description identifies three convolution-pooling stages and two dense layers.Source: §5 · p. 5, Table 1 · p. 5
• Training used supervised RMSprop with backpropagation, mini-batches of 32, a 10^-4 starting learning rate, and 100 iterations.The protocol used ρ = 0.9, ϵ = 1e −7, and categorical cross entropy loss.Source: §5 · p. 5

6 Results

• Performance evaluation used separate training, testing, and validation sets in an 80:10:10 split for each cancer type.Because LC25000 provided no patient information, evaluation was performed at the image level rather than the patient level.Source: §6 · p. 5
• The authors report that both models learn convolutional feature maps whose filters can be visualized and resemble Gabor filters.Figure 3 shows filters from the first, second, and third convolution layers for the colon and lung models.Source: §6 · p. 5, Figure 3 · p. 6
• Accuracy and loss curves show jitters that the authors attribute to dropout, which they say helps the network generalize.Figure 4 plots loss and accuracy over 100 steps for both colon and lung models.Source: §6 · p. 5, Figure 4 · p. 6
• Both models took around 20 epochs to converge, while the colon model had large accuracy and loss spikes around epochs 58 and 80.Its validation loss initially increased through five epochs before beginning to converge.Source: §6 · p. 6
• Table 2 reports training and validation performance metrics for both models at the image level.The supplied table caption identifies its scope as performance metrics for all models.Source: §6 · p. 7, Table 2 · p. 7

7 Conclusions

• On LC25000, the authors report improved CNN accuracy over traditional machine-learning models and transfer-learning deep CNN models using state-of-the-art texture descriptors.The conclusion does not provide the specific accuracy values for these comparisons.Source: §7 · p. 7
• A shallow CNN designed for color-object classification was adapted to lung and colon histopathological images using a strategy that handles high-resolution textured images without reducing their resolution.Source: §7 · p. 7
• Future work proposed exploring other CNN architectures, hyperparameter optimization, neural style transfer for interclass images, and generative models for histopathology images.The proposed generative-model applications include visualizing and exploring mutations across different ontologies.Source: §7 · p. 7