Cocoa diseases constitute a critical threat to global chocolate production, affecting the livelihoods of farmers and the cocoa industry. Among these diseases, Moniliophthora Roreri, commonly known as frosty pod rot, is one of the most devastating, causing cocoa yield losses of up to 40%. Early detection of this disease is essential for timely intervention and minimizing its economic impact. This study explores the use of Machine Learning (ML) algorithms for the detection of Moniliophthora Roreri in cocoa fruits, focusing on the automatic classification of its four developmental stages: humps, oily spots, brown spots, and white powder or sporulation. The focus was on building and training neural networks to detect visual symptoms of cocoa disease. Data from field images of infected and healthy cocoa pods were collected, processed, and analyzed to improve model performance. The study evaluated two neural network architectures, MobileNetV2 and Xception, to assess their effectiveness and reliability in classifying diseased and non-diseased cocoa fruits. The results showed that the MobileNetV2 achieved better performance than Xception, reaching an accuracy of 0.94 and a macro-averaged F1-Score of 0.95, demonstrating their potential for real-world application on cocoa farms.