Mastitis is a disease that is considered an obstacle in dairy farming. Some methods of diagnosing mastitis have been used effectively over the years, but with an associated relative cost that reduces the producer’s profit. In this context, this sector needs tools that offer an early, safe, and non-invasive diagnosis and that direct the producer to apply resources to confirm the clinical picture, minimizing the cost of monitoring the herd. The objective of this study was to develop a predictive methodology based on sequential knowledge transfer for the automatic detection of bovine subclinical mastitis using computer vision. The image bank used in this research consisted of 165 images, each with a resolution of 360 × 360 pixels, sourced from a database of 55 animals diagnosed with subclinical mastitis, all of which were not exhibiting clinical symptoms at the time of imaging. The images utilized in the sequential learning transfer were those of MammoTherm, which is used for the detection of breast cancer in women. The optimized model demonstrated the most optimal network performance, achieving 92.1% accuracy, in comparison to the model with manual search (86.1%). The proposed predictive methodologies, based on knowledge transfer, were effective in accurately classifying the images. This significantly enhanced the automatic detection of both healthy animals and those diagnosed with subclinical mastitis using thermal images of the udders of dairy cows.

The detrimental impact of noisy labels on the generalization performance of deep neural networks has sparked research interest in learning with noisy labels (LNL). Among the various methods proposed to mitigate this detrimental effect, the Co-Teaching method, characterized by co-training with the small-loss criterion, is one of the most established approaches, being widely employed as an important component in recent LNL methods. Although Co-Teaching can mitigate the overfitting effect, it still persists, especially in scenarios with high rates of label noise in datasets. Strategies from the LNL literature to address this typically include the use of disagreement techniques and alternative loss functions. In this paper, we propose the Cyclic Co-Teaching (CCT) method, which employs cyclic variations in the learning rate and sample retention rate at the mini-batch level, along with a checkpoint mechanism that ensures training in subsequent cycles always resumes from the best models obtained so far. For method optimization, we developed a framework that incorporates a pre-training phase to obtain an optimized vanilla model used to initialize CCT model weights, and a transparent univariate optimization strategy for hyperparameters that does not strictly require a clean validation set. Experimental results on synthetic and real-world datasets, under different types and levels of noise and employing various neural network architectures, demonstrate that CCT outperforms several state-of-the-art LNL methods in most evaluated scenarios.