![LungAI.webp](LungAI.webp) # 1. Abstract --- LungAI is an AI-powered platform designed to assist healthcare professionals in the early detection of lung cancer through the analysis of CT scan images. Using deep learning models like ResNet50, it classifies images with 90% accuracy and integrates techniques such as data augmentation and image preprocessing. Additionally, LungAI incorporates Grad-CAM, a tool for generating heatmaps that highlight areas of interest, aiding doctors in interpreting model predictions and identifying potential tumors. The platform also allows for the generation of detailed medical reports, which can be personalized and downloaded. Serving both educational and diagnostic purposes, LungAI demonstrates the potential of AI in healthcare, helping to optimize resources, reduce diagnostic delays, and ultimately improve early cancer detection and patient outcomes. ![Captura de pantalla 2024-12-30 a las 21.44.32.png](Captura%20de%20pantalla%202024-12-30%20a%20las%2021.44.32.png) # 2. Introduction --- According to data from the Spanish Cancer Registry (REDECAN) and the National Statistics Institute (INE), Lung cancer remains the leading causes of death not only in Spain but worldwide. In 2022, lung cancer was the most frequently diagnosed cancer, with 2.5 million new cases globally, and it also had the highest mortality rate, with approximately 1.8 million deaths. In Spain, more than 70% of lung cancer cases are diagnosed at an advanced stage, which significantly reduces the chances of successful treatment. This is where my project, LungAI, comes in. LungAI is a predictive model that aims to identify high-risk patients more efficiently and with high accuracy. It enhances the speed, cost, and accessibility of lung cancer diagnosis, ultimately contributing to earlier detection and better patient outcomes. ![Captura de pantalla 2024-12-30 a las 21.40.39.png](Captura%20de%20pantalla%202024-12-30%20a%20las%2021.40.39.png) # 3. Methodology --- The methodology for developing LungAI involved the integration of advanced machine learning techniques, including deep learning and data augmentation, to create a model capable of accurately identifying lung cancer in CT scan images. This section describes the approach taken in detail, including the preprocessing of data, the architecture of the model, the training process, and the evaluation methods used. ## 3.1 Data Collection and Preprocessing The dataset used for training, validation, and testing consists of CT scan images. These images were preprocessed to ensure compatibility with the model architecture and to improve the model's performance. The images were resized to **224x224** pixels to match the input size required by the pretrained ResNet50 model, which has an input size of 224x224 RGB images. In addition to resizing, the pixel values of the images were **normalized** to a range between 0 and 1 by dividing the pixel values by 255. This normalization step is important because it standardizes the input data, ensuring that the model does not focus on the raw pixel values but instead learns the features that matter. ## 3.2 Data Augmentation To prevent overfitting and make the model more robust, data augmentation was applied to the training set. This technique artificially increases the size of the dataset by generating variations of the images during training. The following augmentation transformations were applied: - Rotation: The images were randomly rotated up to **20 degrees** to simulate different orientations of the lung scans. - Shifting: Horizontal and vertical shifts were applied, randomly moving the images up to 20% of their width and height to account for potential misalignment of the images. - Shearing: A shearing transformation was applied with a range of **0.2**, which distorts the images by slanting them, helping the model generalize better. - Zooming: Random zoom transformations were applied to simulate different levels of magnification, with a zoom range of **0.2**. - Flipping: The images were randomly flipped horizontally, increasing the model’s ability to generalize by simulating images taken from different angles. The **validation** and test sets did not undergo data augmentation, ensuring that the evaluation of the model was based on original, unaltered data. Only image normalization was applied to these datasets. ```python train_datagen = ImageDataGenerator( rescale=1./255, rotation_range=20, width_shift_range=0.2, height_shift_range=0.2, shear_range=0.2, zoom_range=0.2, horizontal_flip=True, fill_mode='nearest' ) valid_test_datagen = ImageDataGenerator(rescale=1./255) train_generator = train_datagen.flow_from_directory( os.path.join(dataset_path, 'train'), target_size=(224, 224), batch_size=32, class_mode='binary' ) valid_generator = valid_test_datagen.flow_from_directory( os.path.join(dataset_path, 'valid'), target_size=(224, 224), batch_size=32, class_mode='binary' ) test_generator = valid_test_datagen.flow_from_directory( os.path.join(dataset_path, 'test'), target_size=(224, 224), batch_size=32, class_mode='binary', shuffle=False ) ``` ## 3.3 Model Architecture The core of **LungAI** is based on the **ResNet50** model, a deep convolutional neural network (CNN) architecture known for its powerful performance in image classification tasks. ResNet50 was initially trained on the **ImageNet** dataset, which contains millions of labeled images from various classes. This pretrained model has learned general features that are applicable across various image domains, making it a suitable choice for transfer learning. **ResNet50** was chosen because of its **skip connections**, which allow the model to skip certain layers and help avoid the vanishing gradient problem in deep neural networks. This architecture is well-suited for tasks that require learning deep representations from complex data like medical images. The model was initialized with **ImageNet weights**, which were kept frozen in the initial layers of the network. The **top layers** of ResNet50 (the fully connected layers responsible for classification) were removed, and custom layers were added for the binary classification task (normal vs. cancerous). **Fine-Tuning** The next step was **fine-tuning** the model to make it specialized for lung cancer detection. Fine-tuning involved unfreezing a portion of the layers in the ResNet50 model to allow the model to adapt the general features learned from ImageNet to the specific characteristics of lung CT scans. The **first 50 layers** of the ResNet50 model were kept frozen, and the layers starting from the **50th layer onward** were unfrozen for training. This approach preserves the general features learned from ImageNet while allowing the model to learn task-specific features related to lung cancer detection. After the ResNet50 base, additional fully connected layers were added to the model: - A **dense layer with 128 units** and **ReLU activation** was added to introduce non-linearity to the model. - Another **dense layer with 256 units** followed, also with **ReLU activation**. - A **dropout layer** with a rate of **0.5** was added to prevent overfitting by randomly setting 50% of the neurons to zero during training. - Finally, the output layer consists of a **single neuron with a sigmoid activation** function, which is typical for binary classification tasks. ## 3.4 Model Compilation The model was compiled using the **Adam optimizer**, a popular optimization algorithm that adapts the learning rate during training. The **learning rate** was set to **0.00005** to ensure stable fine-tuning, preventing the model from updating weights too drastically in the early stages of training. The loss function chosen was **binary cross-entropy**, which is standard for binary classification tasks. The performance metric used to evaluate the model was **accuracy**, which indicates how well the model is classifying images into the correct categories (normal vs. cancerous). ```python model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.00005), loss='binary_crossentropy', metrics=['accuracy']) ``` ## 3.5 Model Training The model was trained using the **ImageDataGenerator** class in Keras, which allows real-time data augmentation during training. The train generator applies the data augmentation transformations to the images on the fly, creating new versions of each image for each epoch. The **validation data** was used to evaluate the model's performance during training, but it was not augmented to ensure that the validation results reflect the model's performance on unaltered images. The model was trained for **20 epochs** with a batch size of **32** images. During training, a **ReduceLROnPlateau** callback was used to decrease the learning rate if the validation loss stopped improving, which helps the model converge faster and avoid overshooting the optimal solution. ```python reduce_lr = ReduceLROnPlateau( monitor='val_loss', factor=0.1, patience=2, min_lr=1e-6 ) model.fit( train_generator, validation_data=valid_generator, epochs=20, callbacks=[reduce_lr], steps_per_epoch=train_generator.samples // train_generator.batch_size, validation_steps=valid_generator.samples // valid_generator.batch_size ) ``` ## 3.6 Model Evaluation After training, the model's performance was evaluated on the **test dataset**. The **test generator** was used to load and preprocess the test data. The final evaluation metrics were **accuracy, ROC AUC score**, and the **confusion matrix**. # 4. Demo video --- Below is a demonstration of the LungAI project:
# 5. Results --- The model was evaluated using the test set, and the performance metrics were calculated to assess its effectiveness. The classification report, which includes precision, recall, and F1-score, is as follows: - **Class 0 (Cancer)**: - Precision: 0.80 - Recall: 1.00 - F1-Score: 0.89 - Support: 177 instances - **Class 1 (Normal)**: - Precision: 1.00 - Recall: 0.77 - F1-Score: 0.87 - Support: 191 instances Overall accuracy of the model on the test set was **88%**. In addition to the per-class performance, the macro and weighted averages are as follows: - **Macro Average**: - Precision: 0.90 - Recall: 0.89 - F1-Score: 0.88 - **Weighted Average**: - Precision: 0.91 - Recall: 0.88 - F1-Score: 0.88 The model's **ROC AUC score**, which evaluates the model's ability to distinguish between the two classes (Cancer vs. Normal), was **0.96**, indicating excellent discriminatory power. ![Captura de pantalla 2024-12-30 a las 21.33.50.png](Captura%20de%20pantalla%202024-12-30%20a%20las%2021.33.50.png) Additionally, a confusion matrix was generated to visualize the model's performance on the test data. The matrix revealed the following results: - **44 false positives**: These are cases where the model predicted the patient to be normal (Class 1), but the patient actually had cancer (Class 0). - **0 false negatives**: This is a critical result, as it indicates that no cancer cases (Class 0) were missed by the model. The model did not incorrectly classify any cancer patient as normal. - **177 true positives**: These are the cases where the model correctly identified cancer (Class 0). - **147 true negatives**: These are the cases where the model correctly identified normal patients (Class 1). It is particularly important that the model achieved **0 false negatives**, meaning that the model did not misclassify any cancer cases (Class 0) as normal (Class 1). This is a crucial outcome, as false negatives can have severe consequences in cancer diagnosis, where early detection is essential for effective treatment. ![Captura de pantalla 2024-12-30 a las 21.30.46.png](Captura%20de%20pantalla%202024-12-30%20a%20las%2021.30.46.png) # 6. Discussion --- ## 6.1. Strengths The LungAI model demonstrates several notable strengths: - **High Accuracy and AUC-ROC Score**: The model achieved an impressive accuracy of 88% and an AUC-ROC score of 0.96, indicating its robustness in distinguishing between patients with and without cancer. - **Zero False Negatives**: A key achievement is the absence of false negatives, meaning the model successfully identifies all cancer cases in the test dataset. This is crucial in clinical settings to ensure no patient with cancer is overlooked. - **Transfer Learning Efficiency**: Leveraging the ResNet50 model pre-trained on ImageNet allowed the model to generalize well to this problem with a relatively small dataset, reducing the time and resources needed for training. - **Flexibility for Explainability**: The model is configured to support Grad-CAM for visual explanations, which can provide additional insights into the regions of interest in the input images. ![Captura de pantalla 2024-12-30 a las 21.44.58.png](Captura%20de%20pantalla%202024-12-30%20a%20las%2021.44.58.png) ## 6.2. Limitations Despite its strengths, the LungAI model has certain limitations: - **Lower Recall for the "Cancer" Class**: Although there are no false negatives, the model's recall for the "cancer" class indicates room for improvement. This could result in a reduced ability to identify all cancer cases in a more diverse dataset. - **Binary Classification Restriction**: Currently, the model performs binary classification (cancer vs. normal). It does not provide information on the specific type of cancer, which could be valuable for tailoring treatment. - **Unresolved Grad-CAM Functionality**: While the model is equipped for Grad-CAM, the implementation remains incomplete. Specifically, the inability to generate colored visualizations of tumor regions in CT images limits its utility in clinical decision-making. ## 6.3. Future Work To address these limitations and enhance the model's capabilities, the following steps are planned: **Improving Recall for the "Cancer" Class** Efforts will be made to reduce false negatives further, even in datasets with more diverse cancer presentations. This may involve fine-tuning hyperparameters, rebalancing the dataset, or employing techniques such as focal loss. **Transition to Multiclass Classification** Developing a more complex model to classify not only the presence of cancer but also its type. Early experiments in this direction have yielded a precision of 40%, indicating a promising area for further exploration. **Enhancing Grad-CAM Visualizations** Debugging and resolving issues with Grad-CAM implementation will be prioritized. This includes identifying why CT scan images fail to display the colored heatmaps indicating tumor regions, ensuring that clinicians can interpret the model's predictions effectively. **Model Optimization** Continued optimization, including experimenting with advanced architectures (e.g., vision transformers) or ensemble methods, could further enhance the model's accuracy and interpretability. ![Captura de pantalla 2024-12-30 a las 21.45.29.png](Captura%20de%20pantalla%202024-12-30%20a%20las%2021.45.29.png) # 7. Conclusion --- LungAI represents a promising advance in early cancer detection, achieving 88% accuracy and a 0.96 AUC-ROC score. Its key strength is zero false negatives, ensuring no cancer cases go undetected. Future improvements include boosting recall, transitioning to multiclass classification, and resolving Grad-CAM issues. These advancements will improve the model's performance and clinical application. This model has the potential to save lives, optimize resources, and reduce healthcare costs. With an initial investment in infrastructure and training, it could be integrated into healthcare systems delivering measurable results. LungAI showcases the transformative potential of AI in medical diagnostics, enabling faster, more accurate cancer detection. # 8. References --- 1. Spanish Cancer Registry (REDECAN) https://redecan.org/es 2. National Statistics Institute (INE) https://www.ine.es/ 3. Spanish Association Against Cancer https://observatorio.contraelcancer.es/ 4.Chest CT-Scan images Dataset https://www.kaggle.com/datasets/mohamedhanyyy/chest-ctscan-images 5. The IQ-OTH/NCCD lung cancer dataset https://www.kaggle.com/datasets/hamdallak/the-iqothnccd-lung-cancer-dataset # 9. Links! --- [My LinkedIn!](https://www.linkedin.com/in/sara-salmon-fernandez-888835175/) [My GitHub! ](https://github.com/SaraSalmon)