Artificial Intelligence in Medicine: Early Diagnosis and the Black Boxes

“Imagine if a cancerous lesion could be detected months before it becomes visible to the doctor’s eye, or if signs of a heart disease could be identified even before the symptoms appear. This is no longer a vision of the future, but today’s reality according to the press release from the project “SustAInLivWork”.

“AI doesn’t get tired or distracted, and it’s results are not influenced by emotional state or time of day,” notes D. Dovydas Verikas, a cardiologist at the Lithuanian University of Health Sciences (LSMU) who uses AI tools in his daily work.

According to him, such consistency gives AI a clear advantage over humans where decisions are related to barely noticeable changes in cardiac or oncological diagnostics.

D. Verikas, a cardiologist, researcher at LSMU, and member of the SustAInLivWork project, which is developing an Artificial Intelligence Excellence Centre in Kaunas, explains that artificial intelligence (AI) uses advanced algorithms to detect subtle changes in medical images the kind that the human eye may overlook or interpret differently depending on the specialist’s experience.

“For example, AI is successfully applied in the analysis of mammograms, computed tomography (CT), magnetic resonance imaging (MRI), ultrasound, and other imaging methods. It helps to identify early signs of cancer, detect neurological disorders, vascular pathologies, and other significant diagnostic markers,” he explains.

D. Verikas nonetheless notes that AI is not meant to replace doctors – it acts as an additional tool that supports a specialist’s expertise and reduces the likelihood of diagnostic errors.

Every Learning Cycle Bringing Increasingly Accurate Diagnoses

In medical image analysis, AI most often relies on deep learning methods, particularly Convolutional Neural Networks (CNNs). These networks are specifically designed for image recognition and are therefore well-suited for interpreting complex medical visuals.

“First, data need to be prepared as the model must be trained on large sets of pre-labelled medical images, for example, X-rays with doctor-assigned diagnoses or marked pathologies. The AI system then analyses these images, learning to recognise specific patterns or structures associated with diseases. Each training cycle enables the algorithm to identify diagnostic signs with greater precision,” the researcher has explained.

The model is subsequently tested on newer, previously unseen images to evaluate its ability to accurately perform in real-world conditions. If systematic inaccuracies are observed, the algorithm is further refined and adapted to new data.

“It’s important to stress that the quality of AI decisions directly depends on the quality of the data – incorrectly labelled or insufficiently diverse data can lead to inaccurate conclusions. That’s why close collaboration between doctors, IT specialists, and data scientists is essential for achieving reliable, practical results,” he notes.

Detecting Minor Structural Changes

According to D. Verikas, AI systems are capable of recognising pathologies due to their ability to process vast amounts of image data and detect extremely fine structural changes that the human eye might miss or misinterpret.

“CNN models can ‘learn’ to identify specific microstructural patterns that doctors often find difficult to spot – such as abnormal cell arrangements, texture changes, small tumour indicators, or calcium deposits in blood vessels. Models like U-Net can precisely segment and localise pathologies, marking their boundaries – for example, tumour edges or areas of ischaemia in cardiac MRI,” he explains.

Furthermore, AI can analyse multi-modal data – meaning it can integrate information from different sources, such as CT and MRI scans, to create a more comprehensive and contextual picture of the patient’s condition.

“At present, AI algorithms work particularly well in medical fields with abundant data and well-defined standards. Excellent results are achieved, for example, in automating standard echocardiographic measurements, identifying breast cancer in mammograms, or detecting lung nodules in CT scans,” the doctor has noted.

They are also successfully applied in areas with well-visualised structures, such as the retina, for detecting signs of diabetic retinopathy. Reliability remains high in areas where clear, standardised diagnostic classifications are used – for instance, assessing echocardiographic or MRI parameters like ejection fraction or chamber dimensions.

AI Does Not Tire, Unlike Humans

When it comes to complex diseases such as early-stage cancer or heart disease, AI has several significant advantages. Firstly, it can process enormous volumes of data that might appear as mere “noise” to the human eye or standard statistical analysis.

“In cardiology, for example, AI can analyse thousands of CT images or echocardiograms and detect subtle changes that a doctor might not yet consider pathological but that may later prove clinically relevant. In this way, AI helps to identify diseases at an early stage when treatment is most effective,” he says.

Secondly, AI tools are exceptionally consistent. Unlike humans, they do not tire, lose focus, or experience emotional fluctuations.

“This means that an algorithm assessing the same CT or echocardiography image will reach the exact same conclusion every time. Such stability is especially valuable in areas like early oncological screening or subtle echocardiographic marker evaluation, where doctors’ opinions may sometimes differ,” D. Verikas adds.

The Final Diagnosis Always Belongs to the Doctor

Nonetheless, it is important to discuss the limitations of this technology. D. Verikas stresses that AI is not a “magic solution” – its performance depends on the data it has been trained on. If the training data represent a limited patient population, the model may perform poorly when applied to other groups. AI can also fail when confronted with unfamiliar cases – such as rare anatomical variations, atypical tumours, or lower-quality images. In such instances, the algorithm may produce an incorrect result while “believing” its decision to be correct.

“Challenges also arise when diagnostic signs lack clear boundaries or depend on disease progression over time. This is common in early stages of Alzheimer’s disease, where an accurate diagnosis requires not just image analysis but also evaluation of disease progression over time,” he explains.

Another limitation is the absence of clinical context. AI can detect early signs of cardiac damage in images, but it cannot assess the patient’s lifestyle, comorbidities, or social factors. The doctor’s role, therefore, is to integrate AI-generated data with the broader context of the patient’s health.

“Another major challenge is the so-called black box problem: it’s not always clear why the algorithm reached a particular decision. In medicine, this is critical, as doctors must understand and explain the basis of a diagnosis to their patients. For this reason, the final decision must always be made by the doctor,” notes the LSMU researcher.

AI Solutions Carefully Tested

One of the main ethical challenges concerns accountability. Patients need to know who is responsible for decisions – the doctor or the algorithm. Data protection is equally important.

“Since AI learns from real medical images, it is essential to ensure that this data is properly anonymised and stored in line with the General Data Protection Regulation (GDPR) and other relevant legislation. Patients must be confident that their information is used responsibly and not for inappropriate purposes,” says the SustAInLivWork project researcher.

AI errors are managed in several ways: algorithms are trained on diverse datasets, continuously tested in real clinical environments, and – most importantly – their results are always verified by specialists. Experience shows that the best results are achieved when AI acts as a “second pair of eyes” for the doctor – helping to spot important details, speeding up analysis, while the doctor ensures the necessary context and takes responsibility for the final diagnosis.

“AI use in medicine is strictly regulated. Before being implemented in clinical practice, all algorithms must meet medical device regulatory standards. In Europe, they must obtain CE marking, and in the United States, FDA approval. This means that every solution undergoes thorough testing to ensure it operates safely, reliably, and truly enhances diagnostic quality,” he explains.

According to D. Verikas, AI holds enormous potential for improving diagnostics, but it must be applied ethically – transparently, safely, and always under medical supervision. Only then can AI become a reliable partner within the healthcare system rather than a potential source of risk.

The SustAInLivWork initiative is the first competence centre of its kind in Lithuania to systematically consolidate AI knowledge and expertise. It brings together four Lithuanian universities – Kaunas University of Technology (KTU), Vytautas Magnus University (VMU), Lithuanian University of Health Sciences (LSMU), and Vilnius Gediminas Technical University (VILNIUS TECH) – in collaboration with partners from Finland (Tampere University) and Germany (Hamburg University of Technology). It is a long-term cross-sectoral platform uniting science, business, the public sector, and society.

Source: lrt.lt