Surgical Robotics Technology

Fujitsu Announces Development of Explainable AI Technology

Fujitsu logo

Fujitsu has announced the development of explainable AI technology that automatically draws on data in multiple formats, including text, images, and numerical data, to create knowledge graphs that will help users to more easily draw meaning from large-scale data sets with high accuracy for areas including medicine.

To confirm the effectiveness of this technology, Fujitsu tested it on several key benchmarks from the medical field, including for lung cancer type classification and breast cancer patient survival prediction. These tests confirmed that Fujitsu’s technology can accurately support the identification of two main types of lung cancer, for example, by illuminating the factors behind the pathological classification based on key visual cues.

Fujitsu has also developed a technology to extract and train algorithms on the distinct features of images with completely different depictions of objects and to make highly accurate judgments. It is anticipated that this technology can be applied to train AI to support highly accurate assessments from pathological images for which sufficient training data cannot be prepared.

Going forward, Fujitsu will continue to develop these multimodal technologies (1) for general use in a range of different fields and disciplines. By the end of fiscal 2024, Fujitsu also plans to offer the newly developed technologies via the Fujitsu Research Portal (2), an environment that gives users the ability to quickly test Fujitsu’s advanced technologies.

Fujitsu AI technology

About the newly developed technology

Fujitsu has been conducting research and development on multimodal technologies that handle data from different fields and in a variety of formats, and has developed the following two AI technologies that combine multiple different data formats and images with completely different depictions of objects to train algorithms and make conclusions from various perspectives.

  1. AI technology that is trained by combining image data with completely different drawing methods, such as line drawings and photographs

Fujitsu has developed a technology that combines data from images in which objects are drawn in completely different ways, such as picture, line drawings, Illustration and photographs, to learn how to accurately differentiate images (for example, to determine what the subject is). With this method, the unique and common feature values of each image type are extracted and used for training, and these unique or common feature values are used for differentiation. This way, even when objects are drawn in different ways the combination of multiple types of data can be used to train the algorithm to make appropriate decisions.

Fujitsu evaluated this technology using three standard benchmarks in this research field (PACS, Office-Home, DomainNet), that consist of datasets of multiple types of images including art, manga images, and photographs. As a result of the benchmark tests, Fujitsu confirmed that with the new technology accuracy of object identification could be improved by about 2% compared with conventional technologies, where objects were identified using image data with completely different drawing methods. This result was recorded in ICLR 2024 of the renowned conference The International Conference on Learning Representations (ICLR) (3), and Fujitsu has presented it on May 8, 2024.

  1. Explainable AI technology that integrates data from different formats and transforms it into a common knowledge graph for training

To integrate different data formats such as not only images but also text and images, Fujitsu has developed a new technology that converts different data into a common knowledge graph regardless of the original format by applying the above technology. This is automatically integrated with AI to create a large-scale integrated knowledge graph, which can be used to allow AI to make decisions in an explainable manner.

By applying this technology to the following medical fields, Fujitsu anticipates outcomes that could exceed the performance of conventional technologies.

In combination with the newly developed technology to extract and train algorithms on the distinct features of images with completely different depictions of objects, the technology is also expected to improve the accuracy of the identification of pathological images for which sufficient training data cannot be prepared.

1) Classification of lung cancer

Treatments are currently being introduced for different types of lung cancer, such as adenocarcinoma (4) and squamous cell carcinoma (5), and accurate classification of cancer types is important to ensure the correct treatment. In the past, a doctor would visually refer to multiple pieces of information and perform a painstaking medical examination. The new technology however offers the potential to significantly streamline this process, as it uses AI to automatically integrate pathological images and genomic information (copy number abnormality information) of lung cancer patients to identify cancer types. As a result, when evaluated using data from The Cancer Genome Atlas (TCGA), a global standard benchmark, the technology achieved the world’s highest accuracy of 92.1% for lung cancer typing, compared with the previous highest accuracy of 87.1%. For these types of classification, the basis of the judgment can be shown by going back to the pathology image data.

2) Determining survival prediction for breast cancer patients Being able to accurately predict the duration of survival for each treatment when a person chooses a treatment method increases the likelihood that the appropriate treatment will be chosen. In this study, by automatically integrating and reviewing breast cancer patient image data as well as RNA data (6) and medical care data using AI, Fujitsu evaluated using benchmark data from The Cancer Genome Atlas (TCGA). In the task of predicting the survival time of breast cancer patients, Fujitsu’s technology achieved 71.8%, compared with the previous best accuracy of 66.8%. Imaging data can be used to provide evidence when predicting these survival times.

Future Plans

The technology will be offered to users via the Fujitsu Research Portal in fiscal 2024.

In 2023, Fujitsu began collaborating with the Barcelona Supercomputing Center – Centro Nacional de Supercomputación (BSC-CNS) (7), the Spanish national supercomputing center, a leading research organization, in the field of personalized medicine. The AI technology developed using this multi-modal technology will also be utilized in joint research with the Barcelona Supercomputing Center to further improve accuracy and gain global recognition. Fujitsu will further develop this technology with a view to utilizing it not only in the medical field, but also in various fields such as data center failure prediction and fraud detection.[1] Multimodal technology :A technology that handles multiple data in different formats such as text, images, and numerical values. In the medical field, it is a technology that can uniformly handle patient information in electronic medical records, test results, CT images, and genomic databases.
[2] Fujitsu Research Portal :Portal site that has been open to the public since June 2023 to provide registered users access to trial versions of Fujitsu’s advanced technologies.
[3] The International Conference on Learning Representations (ICLR) :World’s top international conference on representation learning
[4] Adenocarcinoma :A type of lung cancer that begins in epithelial tissue called glandular tissue. It is the most common type of lung cancer and accounts for about half of the total.
[5] Squamous cell carcinoma :A type of lung cancer that begins in the mucosal tissues of the body called squamous epithelium. It is the second most common type of lung cancer after adenocarcinoma, accounting for about 20 to 30% of the total.
[6] RNA Data :RNA is the source material of proteins that make up an organism, and RNA data is data that can be processed computationally by a sequencer.
[7] Barcelona Supercomputing Center – Centro Nacional de Supercomputación (BSC-CNS) :Location Barcelona, Spain. Director Mateo Valero

Source: Fujitsu

Surgical Robotics Technology

Join thousands of Surgical Robotics Experts and get the latest updates straight to your inbox!