Mafalda Falcão Ferreira

Abstract
Objectives: Artificial intelligence (AI) is poised to transform breast cancer care. However, most scientists, engineers, and clinicians are not prepared to contribute to the AI revolution in healthcare. In this paper, we describe our experiences teaching a new undergraduate course for American students that aims to prepare the next generation for cross-cultural designthinking, which we believe is crucial for AI to achieve its full potential in breast cancer care. Materials and methods: The key course activities are planning, conducting, and interpreting interviews of healthcare professionals from both Portugal and the United States. Since the course is offered as a short-term faculty-led study abroad program in Portugal, students are able to explore the impact of culture on healthcare delivery and the design of healthcare technologies. Results: The learning assessments demonstrated student growth in several areas pertinent for future development of AI for breast cancer care. With respect to understanding breast cancer care, prior to taking this course, most students had underestimated the impact of cancer and its treatment on women's quality of life and most were unaware of the importance of multidisciplinary care teams. Regarding AI in medicine, students became more mindful of data privacy issues and the need to consider the effect of AI on healthcare professionals. Conclusion: This course illustrates the potential benefits for AI in medicine of introducing future scientists, engineers, and clinicians to cross cultural design-thinking early in their educational experiences. (C) 2020 The Author(s). Published by Elsevier Ltd.

Abstract
The importance of early detection of diseases with high-mortality is crucial to save lives. Deep Learning algorithms are recurrently used by many researchers that aim to model the progression and treatment of these conditions. There is growing evidence that the complexity of a Deep Learning model is correlated to its performance: the deeper the network, the more accurate it is. However, as the topology deepens, training gets more demanding: (1) increased need of data, (2) increased computational costs, and (3) increased time for evaluation, fine-tuning, and subsequent feedback-based activities inherent to Data Science, with direct impact on the exploration towards finding the best model, due to an inherent trial-and-error approach. We hypothesize that there exist (domain-specific) architectural patterns that, if applied during the model exploration phase, allow an overall improvement of the training performance. Should it be true, it would significantly reduce the exploration phase length, contributing to both Medicine and Computer Science fields.

Abstract
Background As of today, cancer is still one of the most prevalent and high-mortality diseases, summing more than 9 million deaths in 2018. This has motivated researchers to study the application of machine learning-based solutions for cancer detection to accelerate its diagnosis and help its prevention. Among several approaches, one is to automatically classify tumor samples through their gene expression analysis. Methods In this work, we aim to distinguish five different types of cancer through RNA-Seq datasets: thyroid, skin, stomach, breast, and lung. To do so, we have adopted a previously described methodology, with which we compare the performance of 3 different autoencoders (AEs) used as a deep neural network weight initialization technique. Our experiments consist in assessing two different approaches when training the classification model - fixing the weights after pre-training the AEs, or allowing fine-tuning of the entire network - and two different strategies for embedding the AEs into the classification network, namely by only importing the encoding layers, or by inserting the complete AE. We then study how varying the number of layers in the first strategy, the AEs latent vector dimension, and the imputation technique in the data preprocessing step impacts the network's overall classification performance. Finally, with the goal of assessing how well does this pipeline generalize, we apply the same methodology to two additional datasets that include features extracted from images of malaria thin blood smears, and breast masses cell nuclei. We also discard the possibility of overfitting by using held-out test sets in the images datasets. Results The methodology attained good overall results for both RNA-Seq and image extracted data. We outperformed the established baseline for all the considered datasets, achieving an average F(1)score of 99.03, 89.95, and 98.84 and an MCC of 0.99, 0.84, and 0.98, for the RNA-Seq (when detecting thyroid cancer), the Malaria, and the Wisconsin Breast Cancer data, respectively. Conclusions We observed that the approach of fine-tuning the weights of the top layers imported from the AE reached higher results, for all the presented experiences, and all the considered datasets. We outperformed all the previous reported results when comparing to the established baselines.

Abstract

Abstract
Cancer is one of the most serious health problems of our time. One approach for automatically classifying tumor samples is to analyze derived molecular information. Previous work by Teixeira et al. compared different methods of Data Oversampling and Feature Reduction, as well as Deep (Stacked) Denoising Autoencoders followed by a shallow layer for classification. In this work, we compare the performance of 6 different types of Autoencoder (AE), combined with two different approaches when training the classification model: (a) fixing the weights, after pretraining an AE, and (b) allowing fine-tuning of the entire network. We also apply two different strategies for embedding the AE into the classification network: (1) by only importing the encoding layers, and (2) by importing the complete AE. Our best result was the combination of unsupervised feature learning through a single-layer Denoising AE, followed by its complete import into the classification network, and subsequent fine-tuning through supervised training, achieving an F1 score of 99.61% +/- 0.54. We conclude that a reconstruction of the input space, combined with a deeper classification network outperforms previous work, without resorting to data augmentation techniques.