From Obstacle to Asset: Transforming AI's Role in Radiology

The vast potential of interpretative AIs is often celebrated in conference talks and research papers. However, a closer look at the actual numbers can be quite sobering: the use of AI in radiology is still struggling to take off. The primary reason? It simply does not deliver enough value to offset the costs and time invested in acquiring the AI.

Kim et al conducted a comprehensive study at a major university medical center in the Netherlands. The results are eye-opening, providing valuable insights into the main challenges and strategies for implementing AI effectively. Their analysis is clear, logical, and outlines a robust path forward.

Study Design

Over three years, they observed the implementation of interpretative AI tools in the radiology department of a large university hospital. They collected data through interviews, meetings, and other methods in a systematic analysis that encompassed 228 hours of material.

Let’s look at the obstacles that AI tools face in clinical implementation.

Key Obstacles Hindering AI Implementation in Radiology

1. The complexity of IT infrastructure, which complicates quick and straightforward AI deployments.
2. The absence of automated data routing and processing, requiring significant manual intervention by radiologists and disrupting workflow efficiency.
3. Time-consuming processes required to deploy AI applications within local infrastructures, alongside the need to deploy numerous AI applications due to their narrow use cases.
4. Inadequate definition of clinical use cases by both users and vendors.
5. Limited understanding among clinicians and managers of how to optimally deploy AI in a clinical setting.

Strategies to Overcome These Challenges: Finding the right Tools!

1. Vendor Neutral AI Platform (VNAI): A platform that orchestrates the use of various AI solutions and standardizes the deployment process.
2. Image Processing Group (IPG): This team, consisting of nine radiographers and two clinicians with engineering backgrounds, was tasked with facilitating the integration of AI applications into daily clinical routines. Additionally, they aimed to evaluate the real-world clinical value of these AI tools.
3. Clinical AI Implementation (CAI) Group: Comprising clinicians, data scientists, and legal and ethical experts, this group's mission was to bridge the gap in knowledge and experience with AI. They promoted understanding and assessed the viability of AI solutions using various communication tools such as blog posts, newsletters, and meetups.

How could these tools fix all of the above mentioned problems?

The Vendor Neutral AI Platform (VNAI) utilizes metadata (method, body region, patient age) to route data and activate the available AI solutions in a defined sequence. This orchestration process works automatically without requiring any manual input. The VNAI then forwards the AI output to the user's front-end systems (e.g., PACS).

This also simplifies legal and data privacy concerns because a standardized legal framework can be implemented. According to the publication, this strategic approach has reduced the time needed to implement a new AI tool from years to months.

From a user experience perspective, the biggest obstacles to using AI are:

1. Non-standardized user interfaces and
2. The need to continuously switch between different user interfaces while reporting.

These factors significantly diminish the perceived benefit of using AI. To address this, the focus was on creating a seamless workflow. The essential components to achieve this are:

1. Collaborating with different vendors (AI, PACS, etc.) to standardize data output and
2. Configuring APIs so that user interfaces for each AI application can be standardized.

This approach has greatly reduced workflow disruptions by eliminating the need to switch between different interfaces, as all results are displayed in the same viewer.

The Image Processing Group was established as a hub for deploying new technology. It consisted of radiographers and clinicians with an engineering background. Their primary task was to analyze radiological images and pre-populate reports for radiologists, ensuring AI tools were applied in a structured and standardized manner following established protocols. A beneficial side effect was that by delegating tasks to radiographers, radiologists experienced a workload reduction—not directly due to technical innovations. The clinicians' second major role was to assess the algorithms' performance on the applied data. They also communicated desired changes to AI vendors and contributed to the generation of a seamless workflow.

The Clinical AI Implementation (CAI) Group developed a process to assess AI applications and make standardized decisions about whether to adopt them. Upon proposing an AI application, its viability and requirements were evaluated by a multidisciplinary group. AI champions were designated in each radiology subspecialty to ensure that the proposed AI solutions were based on realistic use cases. When a proposal was accepted, a project team consisting of a clinician and an IT engineer was formed to further evaluate and implement the idea. Involving clinicians from start to finish ensured that:

1. The AI solution addressed real-life clinical problems, and
2. The radiologists’ experiences with the AI solution were realistic.

The results of this holistic approach

A holistic approach restructures and reconsiders an entire workflow and workforce. While initially time-consuming and labor-intensive, it offers significant long-term benefits: enduring value and enhanced efficiency. In contrast, simply adding an AI tool to existing structures and workflows does not result in long-term positive impacts.

To successfully implement a holistic approach, several key elements are necessary:

1. Collaboration and knowledge sharing across the institute,
2. Centralization of efforts and control,
3. Investments in AI for standardization and automation,
4. Inclusion of users throughout the entire process.

My take on this

The holistic approach is undoubtedly the way to go. There is an abundance of non-holistic approaches, and from what I understand, none have proven effective. At least, I have yet to hear of an AI-enhanced radiology department that truly embodies this title. The plan detailed in this publication is comprehensive, but the so-called 'last mile' is still not fully addressed. The critical question remains: When do we review the AI results, and do we use them as a starting point or a refinement? This likely depends on the specific use case. An interesting publication by Wenderott et al explores this question in detail. Here are their suggested workflows for the last mile:

So, what do you do? Do you look at the AI results first or check them after you’ve completed your report?

Sources

1. Kim, B., Romeijn, S., van Buchem, M. et al. A holistic approach to implementing artificial intelligence in radiology. Insights Imaging 15, 22 (2024). https://doi.org/10.1186/s13244-023-01586-4
2. Wenderott K, Krups J, Luetkens JA, Weigl M. Radiologists' perspectives on the workflow integration of an artificial intelligence-based computer-aided detection system: A qualitative study. Appl Ergon. 2024 May;117:104243. doi: 10.1016/j.apergo.2024.104243. Epub 2024 Feb 1. PMID: 38306741.