Resumen de Deep learning strategies for overcoming diagnosis challenges with limited annotations
María Rocío del Amor del Amor
In recent years, deep learning (DL) has become one of the main areas of artificial intelligence (AI), driven mainly by the advancement in processing power. DL-based algorithms have achieved amazing results in understanding and manipulating various types of data, including images, speech signals and text.
The digital revolution in the healthcare sector has enabled the generation of new databases, facilitating the implementation of DL models under the supervised learning paradigm. Incorporating these methods promises to improve and automate the detection and diagnosis of diseases, allowing the prediction of their evolution and facilitating the application of clinical interventions with higher efficacy.
One of the main limitations in the application of supervised DL algorithms is the need for large databases annotated by experts, which is a major barrier in the medical field. To overcome this problem, a new field of developing unsupervised or weakly supervised learning strategies using the available unannotated or weakly annotated data is opening up. These approaches make the best use of existing data and overcome the limitations of reliance on precise annotations.
To demonstrate that weakly supervised learning can offer optimal solutions, this thesis has focused on developing different paradigms that allow training models with weakly annotated or non-expert annotated databases. In this regard, two data modalities widely used in the literature to study various types of cancer and inflammatory diseases have been used: omics data and histological images. In the study on omics data, methods based on deep clustering have been developed to deal with the high dimensions inherent to this type of data, developing a predictive model without requiring annotations.
In comparison, the results of the proposed method outperform other existing clustering methods.
Regarding histological imaging studies, the detection of different diseases has been addressed in this thesis, including skin cancer (spitzoid melanoma and spindle cell neoplasms) and ulcerative colitis. In this context, the multiple instance learning (MIL) paradigm has been employed as the baseline in all developed frameworks to deal with the large size of histological images.
Furthermore, diverse learning methodologies have been implemented, tailored to the specific problems being addressed. For the detection of spitzoid melanoma, an inductive learning approach has been used, which requires a smaller volume of annotations. To address the diagnosis of ulcerative colitis, which involves the identification of neutrophils as biomarkers, a constraint learning approach has been utilized. With this method, the annotation cost has been significantly reduced while achieving substantial improvements in the obtained results. Finally, considering the limited number of experts in the field of spindle cell neoplasms, a novel annotation protocol for non-experts has been designed and validated. In this context, deep learning models that work with the uncertainty associated with such annotations have been developed.
In conclusion, this thesis has developed cutting-edge techniques to address the medical sector's challenge of precise data annotation. Using weakly annotated or non-expert annotated data, novel paradigms and methodologies based on deep learning have been proposed to tackle disease detection and diagnosis in omics data and histological images. These innovations can improve effectiveness and automation in early disease detection and monitoring.
