Metis-Project-Four

Metastatic Tumor Detection

This project used data consisting of images that showed lymph node tissues under a microscope. They were labelled with tumor or no tumor tissue present in the image. Tumor detection in lymph nodes is especially important when determining the extent of metastasis in a person's cancer diagnosis. This categorization will shape the treatment and recovery of the patient depending on the analysis from a lymph node. The project deployed two separate convolutional neural networks (CNN)s. One was build and then iterated through to the final version and then compared to the second that used transfer learning (ResNet50). Overall, the project sought to aid physicians when determining which lymph's to biopsy for further analysis and detect pixel sized tumor tissue present in otherwise healthy images.

This project was conducted on Google Colab which employs high-Ram and GPUs capable of computing the analysis with minutes that would normally take longer to calculate. The analysis makes extensive use of tensorflow, which as of March 2021 is not compatible with MacOS M1 chips.

Files

- Database Build and Analysis :

• The data consists of 220,000 microscopic slide images of lymph node tissue from the Netherlands in 2016. The dataset was rebalanced from the original to exclude duplicates reducing the original 300,000+ down to the 220,000 used for this analysis made available on Kaggle. The data has an imbalance between Cancerous and Non-cancerous slides of roughly 46% to 54% respectively. The project employed under-sampling of 89,000 for each category setting the remaining images for testing against the models training and validation datasets.

- Baseline & Final Iteration Convolutional Neural Networks:

• Baseline model deployed three layers of convolutional Neural Networks with a binary class mode. This yielded decent accuracy results for the training set, but failed to improve the validation accuracy beyond the 50 ~ 53% range. The model overfit the data and could not be deployable beyond this experiment.
• The Final Iteration closed the gap between the validation accuracy and the training accuracy, but a small gap persisted. This indicated that despite improving the validation accuracy from 50 to 84 that the training data still overfit the model slightly. However when compared to the transfer learning example we see the model outperformed the ResNet50 pre-trained model by roughly 10 points.
• ResNet50 is a Residual Neural Network that is pre-trained on data externally. It relies on those images to build the CNN, by detecting curves, lines, patterns to build shapes and then identify the

Conclusions

• The comparison between the Transfer Learning ResNet50 CNN and the Final Iteration of our model revealed that our model had a higher accuracy in both validation and training. However, the model is still overfitting on the scores when compared to each other internally. The results for the final model were .87 training accuracy and .84 validation accuracy compared to the ResNet50 model at .76 training and validation accuracies each.