With the increasing adoption of teledermatology, there is a need to improve the automatic organization of medical records, being dermatological image modality a key filter in this process. Although there has been considerable effort in the classification of medical imaging modalities, this has not been in the field of dermatology. Moreover, as various devices are used in teledermatological consultations, image acquisition conditions may differ. In this work, two models (VGG-16 and MobileNetV2) were used to classify dermatological images from the Portuguese National Health System according to their modality. Afterwards, four incremental learning strategies were applied to these models, namely naive, elastic weight consolidation, averaged gradient episodic memory, and experience replay, enabling their adaptation to new conditions while preserving previously acquired knowledge. The evaluation considered catastrophic forgetting, accuracy, and computational cost. The MobileNetV2 trained with the experience replay strategy, with 500 images in memory, achieved a global accuracy of 86.04% with only 0.0344 of forgetting, which is 6.98% less than the second-best strategy. Regarding efficiency, this strategy took 56 s per epoch longer than the baseline and required, on average, 4554 megabytes of RAM during training. Promising results were achieved, proving the effectiveness of the proposed approach.

Teledermatology has developed rapidly in recent years and is nowadays an essential tool for early diagnosis. In this work, we aim to improve existing Teledermatology processes for skin lesion diagnosis by developing a deep learning approach for risk prioritization with a dataset of retrospective data from referral requests of the Portuguese National Health System. Given the high complexity of this task, we propose a new prioritization pipeline guided and inspired by domain knowledge. We explored automatic lesion segmentation and tested different learning schemes, namely hierarchical classification and curriculum learning approaches, optionally including additional patient metadata. The final priority level prediction can then be obtained by combining predicted diagnosis and a baseline priority level accounting for explicit expert knowledge. In both the differential diagnosis and prioritization branches, lesion segmentation with 30% tolerance for contextual information was shown to improve classification when compared with a flat baseline model trained on original images; furthermore, the addition of patient information was not beneficial for most experiments. Curriculum learning delivered better results than a flat or hierarchical approach. The combination of diagnosis information and a knowledge map, created in collaboration with dermatologists, together with the priority achieved interesting results (best macro F1 of 43.93% for a validated test set), paving the way for new data-centric and knowledge-driven approaches.

Ceramics is a millenary industrial sector with relevant financial impact for several countries. Efficiency in color mixing is crucial in the ceramic industry, both in terms of staff time and consumables costs. Traditional color mixing methods usually consist of manual processes based on in-depth domain knowledge of basic color theory or color models like Kubelka-Munk. Thus, the efficiency of these procedures is highly dependent on the technician’s expertise, being challenging for novices to acquire these skills and be proficient. This work explores the usage of Deep Learning to generate color mixture predictions in ceramic glazes. The proposed solution is based on spectral data of ceramic components (pigments and glazes) and, based on their respective quantities, simulates the color mixing result in a wholly digital way. Given the lack of freely available datasets, we started by exploring our approach in the NTU Watercolor Pigments Spectral Measurement dataset. We then translated the collected knowledge to the Matcerâmica Ceramics Spectral Measurement dataset, which was specifically created in the ambit of this work. By using data-centric optimization techniques to improve our model interactively, the best performance was achieved by fully connected neural network model, with mean of 1.57 and 2.35 for the NTU and Matceramica datasets, respectively. These results demonstrate the potential of the proposed approach to be integrated into an AI-powered software solution that improves color mixing procedures in the ceramic industry.

Quality control procedures in the manufacturing of tableware ceramics require a demanding, monotonous, subjective, and faulty human manual inspection. This paper presents two machine learning strategies and the results of a semi-automated visual inspection of ceramics tableware applied to a private dataset acquired during the VAICeramics project. In one method, an anomaly detection step was integrated to pre-select possible defective patches before passing through an object detector and defects classifier. In the alternative one, all patches are directly provided to the object detector and then go through the classification phase. Contrary to expectations, the inclusion of the anomaly detector demonstrated a slight reduction in the performance of the pipeline, which may result from error propagation. Regarding the proposed methodology for defect detection, it exhibits average performance in monochromatic images with more than 600 real defects in total, efficiently identifying the most common defect classes in highly reflective surfaces. However, when applied to newly acquired images, the pipeline encounters challenges revealing a lack of generalization ability and experiencing limitations in detecting specific defect classes, due to their appearance and limited available samples used for training. Only two defect types presented high classification performance, namely Dots and Cracked defects.

The ceramic industry is a highly competitive millenary sector with a substantial economic impact in several countries translated into a high global volume business, especially in exports. Reliability in the color development process is critical, ensuring that the produced pieces achieve the defined requirements. Nevertheless, current strategies for color formulation or adjustment are quite manual and subjective, essentially based on a trial-error process. These conventional procedures lead to the development of unnecessary pieces until the target color is achieved, which translates into a waste of raw materials and working time. In this paper, we present an automated approach based on Reinforcement Learning (RL) to improve color adjustments in the ceramic industry. By using the spectral data of the available components (pigments and glazes), the proposed algorithm provides the best formula to achieve the target color, i.e. the list of components for the mixture and corresponding quantities. Two datasets were used: the NTU Watercolor Pigments Spectral Measurement dataset; and the Matcerâmica Ceramics Spectral Measurement dataset. Three RL algorithms were trained and compared for benchmarking purposes: Deep Q-learning. (DQN), Advantage Actor-Critic (A2C) and Continual RL Without Conflict (OWL). The A2C and OWL models obtained similar performances for the NTU dataset with a mean of 0.668 and 0.733, respectively. For the Matcerâmica dataset, OWL yielded better results with a mean of 3.258. These results demonstrate the potential of the proposed approach to be integrated into an AI-powered software solution that optimizes the iterative process of color (re)creation in ceramic glazes.