How Artificial Intelligence Could Change Diagnosis, Treatment and Surgery in Urology

By Mr Neil Haldar, Consultant Urological Surgeon

Artificial intelligence is beginning to influence almost every stage of modern medicine.

Urology may prove to be one of the specialties in which its impact is felt earliest.

That is partly because so much of contemporary urology depends upon information that computers can analyse: MRI and CT scans, microscopic images of prostate biopsies, PSA results, molecular tests, robotic surgery video and large datasets containing the outcomes of thousands of patients.

AI can potentially identify patterns within this information that are difficult—or sometimes impossible—for a human clinician to recognise consistently.

We are therefore moving beyond the idea of artificial intelligence simply “reading a scan”.

AI is beginning to be investigated or used to:

• identify suspicious prostate lesions on MRI;
• detect cancer within digital prostate-biopsy slides;
• estimate the aggressiveness of prostate cancer;
• predict which patients are most likely to benefit from particular treatments;
• highlight bladder tumours during cystoscopy;
• analyse kidney tumours on CT;
• assess the technical performance of robotic surgeons;
• help doctors interpret clinical guidelines;
• produce patient information;
• summarise consultations;
• and reduce the substantial administrative burden associated with modern clinical practice.

But there is an important distinction.

AI in urology is not a single technology at a single stage of development.

Some applications are already regulated and being used clinically. Others have impressive research results but remain experimental. Some of the most ambitious ideas—such as autonomous robotic surgery—remain much further away.

The central question, then, is not whether AI is coming to urology.

It already has.

The more important question is how we use it safely, intelligently, and in a way that genuinely improves patient care.

In brief: how is AI being used in urology?

The most advanced applications of artificial intelligence in urology currently involve:

Prostate MRI – helping radiologists identify and characterise suspicious lesions.

Digital pathology – analysing prostate-biopsy slides and highlighting areas suspicious for cancer.

Prostate cancer prediction – combining pathology and clinical information to estimate prognosis and potentially predict treatment benefit.

Bladder cancer detection – analysing cystoscopy video and highlighting suspicious areas.

Kidney cancer imaging – helping characterise renal masses and predict tumour behaviour.

Robotic surgery – analysing surgical video and instrument movements, assisting navigation and measuring surgical performance.

Large language models – helping with patient information, guideline interpretation, documentation and administrative tasks.

For the foreseeable future, these technologies are more likely to augment clinicians than replace them.

Why is urology particularly suited to artificial intelligence?

Many areas of urology already generate highly structured digital information.

Consider a man being investigated for possible prostate cancer.

His pathway might generate:

• PSA results;
• prostate volume measurements;
• MRI images;
• a PI-RADS assessment;
• biopsy histology;
• Gleason or Grade Group information;
• genomic or molecular information;
• PSMA PET imaging;
• treatment details;
• PSA follow-up;
• continence outcomes;
• erectile-function outcomes.

For thousands of men treated over many years, this becomes an enormous dataset.

Traditional medicine has dealt with this by grouping patients into broad categories such as low-, intermediate- and high-risk disease.

AI makes something more individualised possible.

Instead of asking:

“What usually happens to patients in this category?”

we may eventually be able to ask:

“What is most likely to happen to this particular patient, given everything we know about his cancer?”

That distinction could become extremely important.

AI and prostate MRI

Multiparametric MRI has transformed the diagnosis of prostate cancer.

Before widespread use of MRI, men with an elevated PSA would commonly proceed directly to prostate biopsy.

MRI now allows us to identify suspicious areas within the prostate before biopsy, meaning biopsies can be targeted more accurately and some men may avoid biopsy altogether.

However, prostate MRI interpretation is not perfect.

Radiologists assess suspicious lesions using the PI-RADS scoring system, but interpretation can vary between readers.

An experienced specialist prostate radiologist may interpret a difficult scan differently from someone who reads relatively few prostate MRIs.

This is precisely the sort of problem AI may help address.

Computer-aided systems can analyse MRI images and:

• identify suspicious regions;
• outline potential lesions;
• estimate the likelihood of clinically significant cancer;
• provide an additional assessment alongside the radiologist;
• and potentially reduce variation between different readers.

The goal should not be to remove the radiologist from the process.

A more realistic model is:

Radiologist + AI rather than radiologist versus AI.

AI may become particularly valuable as a second reader—highlighting areas that warrant closer inspection while the radiologist retains responsibility for interpreting the scan in the context of the whole patient.

Could AI decide whether a man needs a prostate biopsy?

Potentially—but this requires greater caution.

A decision about biopsy may depend upon:

• PSA;
• PSA density;
• MRI findings;
• prostate size;
• age;
• family history;
• ethnicity;
• previous biopsy results;
• medications;
• general health;
• and the patient’s personal attitude towards risk.

AI systems may eventually integrate all of these variables and calculate an individual probability of clinically significant cancer.

That could help reduce unnecessary biopsies while ensuring that important cancers are not missed.

However, missing a clinically significant prostate cancer has very different consequences from performing an unnecessary biopsy.

Any system used to influence this decision therefore requires rigorous external validation.

AI and prostate biopsy pathology

One of the most mature AI applications in urology is digital pathology.

Traditionally, pathologists place prostate-biopsy tissue on glass slides and examine it under a microscope.

Increasingly, these slides can be digitised at very high resolution.

Once pathology becomes digital, AI can analyse the image much like computer-vision systems analyse photographs.

Algorithms can help:

• highlight areas suspicious for prostate cancer;
• identify very small tumour deposits;
• estimate tumour volume;
• assess Gleason patterns;
• quantify different components of the tumour;
• and direct the pathologist’s attention towards unusual areas.

This could have several advantages.

Pathologists spend less time examining clearly benign regions and more time evaluating the areas most likely to matter.

AI may also help improve consistency between pathologists when assigning tumour grade.

Once again, however, the most credible model is not an AI system issuing a cancer diagnosis independently.

It is a pathologist supported by computational tools that make the examination more efficient and reproducible.

The bigger opportunity: can AI predict how a prostate cancer will behave?

This may ultimately prove more important than automated cancer detection.

For decades, prostate-cancer treatment decisions have depended upon relatively simple clinical information:

• PSA;
• tumour stage;
• Grade Group;
• number of positive biopsies;
• extent of tumour within each biopsy;
• imaging;
• and increasingly molecular tests.

We then place patients into broad risk categories.

But two cancers that look similar under the microscope may behave very differently.

AI can analyse subtle patterns within tissue that are invisible to the human eye.

Modern multimodal systems can combine digital pathology with clinical information to generate individualised predictions.

This is known as multimodal artificial intelligence.

Instead of asking an algorithm only to examine the biopsy, we can provide it with:

• the biopsy image;
• PSA;
• age;
• tumour stage;
• and potentially imaging, genomic and outcome information.

The system can then look for combinations of features associated with what happened to similar patients in the past.

Can AI predict which prostate-cancer treatment will work?

This is one of the most exciting developments in the field.

Most clinical prediction tools are prognostic.

They answer questions such as:

How likely is this cancer to recur?

A truly useful treatment biomarker needs to go further.

It should be predictive.

It should help answer:

Is this particular patient more likely to benefit from adding this particular treatment?

That distinction matters enormously.

Imagine two men who appear to have similar prostate cancers.

Both might traditionally be offered radiotherapy plus hormone treatment.

An AI-derived biomarker might eventually show that:

• one man has a substantial predicted benefit from additional hormone therapy;
• while the other has only a very small predicted benefit.

If validated prospectively, this could make treatment more individualised.

The objective would not simply be to treat cancer more aggressively.

It would also be to avoid unnecessary treatment.

That may mean:

• avoiding unnecessary hormone therapy;
• avoiding unnecessary radiotherapy;
• identifying patients who require treatment intensification;
• and reassuring patients who can safely receive less treatment.

In my view, this represents one of the most clinically important potential uses of artificial intelligence in prostate cancer.

From risk groups towards individual risk

This illustrates a wider change that AI may bring to medicine.

Conventional medicine often works by placing patients into groups.

For example:

• low risk;
• intermediate risk;
• high risk.

AI may gradually move us towards genuinely individual predictions.

A future prostate-cancer model might incorporate:

• PSA trajectory;
• MRI;
• PSMA PET;
• digital pathology;
• tumour genetics;
• inherited genetic risk;
• previous treatment outcomes;
• age;
• other medical conditions;
• and potentially thousands of subtle imaging and tissue features.

Instead of saying:

“Men in your category have approximately this risk…”

we may increasingly be able to say:

“Based on the characteristics of your cancer, your predicted benefit from treatment A is X, compared with Y from treatment B.”

That would profoundly change shared decision-making.

AI and bladder cancer

Artificial intelligence is also being investigated in bladder-cancer diagnosis.

Bladder cancer is usually detected using cystoscopy—a small flexible camera passed through the urethra into the bladder.

Experienced urologists are extremely good at recognising typical papillary bladder tumours.

The greater challenge can be:

• very small lesions;
• subtle areas of abnormal mucosa;
• flat carcinoma in situ;
• inflammatory changes that mimic cancer;
• and lesions hidden within a complex bladder.

Computer-vision systems can analyse cystoscopy video in real time.

The principle is similar to driver-assistance technology in a car.

The urologist still examines the bladder, but the software may place a box, outline, or warning over an area it considers suspicious.

This could potentially:

• reduce missed lesions;
• improve consistency;
• help train younger urologists;
• and provide an objective record of what was seen.

Research results are encouraging, but AI cystoscopy is earlier in clinical development than some prostate pathology applications.

The major challenge is ensuring that systems perform reliably across:

• different cystoscopes;
• different hospitals;
• different patient populations;
• bleeding;
• inflammation;
• previous treatment changes;
• and unusual tumours.

A system that performs extremely well on the images on which it was developed may perform less well in another hospital.

That problem—generalisation—is one of the central challenges for medical AI.

AI and kidney tumours

Small kidney tumours present another attractive application.

Modern CT and MRI scans frequently identify small renal masses incidentally.

Not all of these masses are malignant.

Some are:

• renal-cell carcinomas;
• oncocytomas;
• angiomyolipomas;
• cystic lesions;
• or other benign abnormalities.

Sometimes imaging cannot confidently distinguish between them.

AI-based radiomics attempts to extract quantitative information from scans that the human eye cannot easily identify.

An algorithm may analyse:

• texture;
• shape;
• contrast enhancement;
• internal heterogeneity;
• the tumour’s relationship with the kidney;
• and potentially hundreds or thousands of other mathematical features.

Future systems could help predict:

• whether a mass is benign or malignant;
• the likely type of kidney cancer;
• tumour aggressiveness;
• suitability for active surveillance;
• and risk of recurrence following surgery.

At present, however, much of renal radiomics remains investigational.

External validation remains essential before such models can replace established diagnostic pathways.

AI and kidney stones

Kidney-stone disease generates another large data source.

AI systems are being studied to predict:

• stone composition from CT;
• likelihood of successful shockwave treatment;
• the probability of becoming stone-free following ureteroscopy;
• outcomes after percutaneous nephrolithotomy;
• and the likelihood of recurrent stone formation.

This may eventually help personalise treatment.

For example, rather than advising that a stone is “probably suitable” for shockwave treatment, a future system might combine:

• stone size;
• position;
• density;
• skin-to-stone distance;
• anatomy;
• patient characteristics;

and produce an individual probability of successful clearance.

Again, this would support rather than replace clinical judgement.

Artificial intelligence and robotic surgery

Robotic surgery is likely to become one of the most visible areas of surgical AI.

Current surgical robots do not perform operations independently.

The surgeon remains in control.

However, robotic systems generate extraordinary amounts of information.

During a robotic prostatectomy, the system potentially records:

• instrument movements;
• instrument speed;
• camera position;
• instrument collisions;
• time spent on each stage;
• surgical video;
• and increasingly information about force applied to tissue.

AI can analyse these data.

That could allow us to measure surgical performance objectively rather than relying entirely upon subjective observation.

Can AI tell whether a surgeon is performing well?

To some extent, this is already possible experimentally.

Algorithms can analyse factors such as:

• economy of movement;
• instrument smoothness;
• unnecessary movement;
• camera control;
• suturing patterns;
• speed;
• and consistency.

Researchers have investigated whether these measurements correlate with outcomes such as urinary continence following robotic prostatectomy.

The long-term potential is substantial.

A surgeon might eventually receive a detailed post-operative analysis showing:

• where movements were efficient;
• where unnecessary traction occurred;
• how performance compared with personal previous cases;
• whether technique differed from high-performing benchmarks;
• and which aspects could be improved.

This could transform surgical training.

Instead of simply saying:

“That dissection could have been better,”

a training system might identify exactly what differed.

Could AI guide a surgeon during an operation?

Potentially.

One of the most exciting developments involves combining AI with three-dimensional imaging and augmented reality.

Before a kidney operation, for example, CT data can be converted into a three-dimensional model showing:

• the kidney;
• tumour;
• arteries;
• veins;
• and collecting system.

A similar model can potentially be generated from prostate MRI showing the position of the tumour.

The difficulty is aligning that virtual model with the real organ during surgery.

Organs move and deform.

AI-based computer vision may eventually let the robotic system continuously recognise anatomical landmarks and keep the digital model aligned with the live surgical image.

A surgeon might then see:

• the estimated position of a tumour below the surface;
• the course of an important blood vessel;
• the relationship between the cancer and a nerve;
• or a warning when the dissection approaches a predefined boundary.

The comparison with satellite navigation is imperfect, but useful.

The surgeon remains responsible for driving.

The system provides a much more sophisticated map.

Will AI eventually perform surgery independently?

AI is likely to perform increasingly complex parts of operations.

That is different from saying that an autonomous robot will replace the surgeon.

Experimental systems have already demonstrated autonomous performance of tightly defined surgical tasks in laboratory and animal settings.

Potential future automated tasks include:

• camera control;
• maintaining tissue retraction;
• placement of sutures;
• image alignment;
• measuring tissue;
• maintaining predefined instrument forces;
• and completing defined components of an anastomosis.

The important distinction is between:

task autonomy

and

complete procedural autonomy.

Performing a repeated series of sutures is fundamentally different from performing an entire cancer operation.

A surgeon continuously deals with uncertainty:

• unexpected anatomy;
• bleeding;
• scar tissue;
• abnormal vessels;
• tumour extent that differs from imaging;
• tissue that behaves unexpectedly;
• and decisions that balance cancer clearance with quality of life.

For the foreseeable future, the most plausible model is surgeon-led surgery with increasing levels of machine assistance and supervised automation.

Large language models and urology

Not all medical AI involves images or robots.

The arrival of large language models has created a completely different category of application.

These systems can process and generate text and may assist with:

• summarising medical information;
• explaining diagnoses;
• comparing treatment options;
• producing patient information;
• drafting clinic letters;
• reviewing clinical guidelines;
• preparing multidisciplinary meetings;
• and answering patient questions.

Research in urology has already shown that general-purpose language models can often produce reasonable answers to clinical questions.

However, they can also give incomplete or incorrect answers with considerable confidence.

This is particularly important in medicine.

An answer that sounds convincing is not necessarily an answer that is correct.

Could an AI use clinical guidelines?

One promising approach is to connect a language model directly to trusted medical guidelines.

Instead of relying primarily on information learned during general training, the system retrieves information from sources such as:

• European Association of Urology guidelines;
• NICE guidance;
• professional consensus documents;
• institutional pathways;
• and peer-reviewed literature.

This is often known as retrieval-augmented generation.

A urology-specific system might eventually allow a clinician to ask:

“What are the guideline-recommended management options for this patient?”

and receive an answer linked directly to the relevant source.

That is potentially much safer than an unconstrained chatbot.

Even then, the system should support, not replace, the clinician.

Guidelines cannot account for every factor affecting an individual patient.

AI and patient information

Large language models could also improve patient communication.

One persistent problem in medicine is that professional guidelines are written for clinicians rather than patients.

AI can potentially translate complex guidance into:

• plain English;
• different reading levels;
• different languages;
• question-and-answer formats;
• summaries;
• and personalised explanations.

Imagine a patient leaving a consultation following a diagnosis of prostate cancer.

With appropriate safeguards, an AI system could generate:

• a personalised summary of the consultation;
• explanations of unfamiliar terminology;
• information about the proposed treatment;
• questions to consider before the next appointment;
• and links to reliable resources.

Used well, this could improve understanding and shared decision-making.

Used badly, it could generate misinformation at enormous scale.

Human oversight remains essential.

The less glamorous AI revolution: reducing medical administration

The area in which AI may change doctors’ working lives most rapidly may have nothing to do with diagnosing cancer.

It is documentation.

Modern clinicians spend a substantial proportion of their working day:

• typing consultation notes;
• generating letters;
• completing referral forms;
• coding diagnoses;
• preparing MDT discussions;
• and entering information into electronic systems.

Ambient AI systems can listen to a consultation—with appropriate consent and data safeguards—and automatically produce a structured clinical note or draft letter.

The clinician then reviews and approves it.

If these systems prove accurate and safe at scale, they could return substantial time to doctors and patients.

The goal should not simply be greater productivity.

It should be more time spent actually listening to patients rather than looking at a computer screen.

What are the risks of AI in urology?

The potential is substantial, but so are the risks.

AI can be confidently wrong

Generative AI can produce inaccurate information that sounds authoritative.

This is often called hallucination.

In medicine, such errors can be dangerous.

Training data can contain bias

If an AI system is developed mainly using data from one population, it may perform worse in another.

Important differences could involve:

• ethnicity;
• age;
• geography;
• socioeconomic background;
• healthcare system;
• and prevalence of disease.

An algorithm that performs well overall can still systematically perform poorly for a particular group.

AI may not explain its reasoning

Complex deep-learning systems may identify patterns without providing an intuitive explanation for why a particular prediction was made.

That becomes especially important if an algorithm recommends one cancer treatment over another.

Clinicians may trust AI too much

A potentially dangerous situation arises when a clinician stops critically evaluating a recommendation because the computer is usually correct.

Automation can reduce human error.

It can also create new forms of human error.

Medical data are extremely sensitive

Advanced AI requires enormous amounts of data.

That raises questions about:

• consent;
• privacy;
• ownership;
• data transfer;
• cybersecurity;
• and commercial use of patient information.

These questions need to be addressed before widespread deployment rather than afterwards.

Could AI create two levels of urological care?

This is an important concern.

The most advanced AI systems may initially be concentrated in:

• major cancer centres;
• academic hospitals;
• well-funded private hospitals;
• and healthcare systems with sophisticated digital infrastructure.

Patients elsewhere may have less access.

That could create a future in which one patient receives:

• AI-enhanced MRI;
• computational pathology;
• personalised molecular prediction;
• AI-assisted robotic surgery;
• and digitally supported follow-up;

while another patient with the same disease receives none of these technologies.

But the reverse is also possible.

Once validated, digital tools can sometimes be distributed more easily than specialist human expertise.

AI could eventually help smaller hospitals by providing:

• expert-level decision support;
• imaging assistance;
• automated second opinions;
• virtual specialist input;
• and more consistent guideline-based care.

Whether AI increases or reduces inequality will depend on how healthcare systems choose to deploy it.

Will AI replace urologists?

I do not think replacement is the most likely outcome.

AI will change what urologists do.

A future urologist may spend less time:

• manually reviewing routine data;
• writing repetitive documentation;
• measuring lesions;
• searching guidelines;
• and completing administrative tasks.

They may spend more time:

• making complex decisions;
• discussing uncertainty;
• interpreting AI-generated predictions;
• performing and supervising procedures;
• managing unexpected findings;
• and helping patients decide what matters to them.

Medicine contains something that AI finds difficult to reproduce:

Judgement under uncertainty.

Often, no single objectively correct treatment exists.

A 58-year-old man with prostate cancer may make a very different decision from an 80-year-old man with exactly the same tumour.

One patient may place enormous importance on avoiding treatment side effects.

Another may prioritise cancer control above almost everything else.

AI can provide increasingly sophisticated risk estimates.

It cannot decide a patient’s priorities.

What might AI in urology look like in ten years?

Predictions should always be treated cautiously, but several developments appear plausible.

1. AI becomes a routine second reader

Radiology and pathology systems may routinely provide an independent computer-generated assessment alongside the clinician.

2. Cancer treatment becomes increasingly personalised

AI models may combine imaging, pathology, genetics and clinical data to estimate the benefit of different treatments for individual patients.

3. The robotic operating theatre becomes intelligent

Robots may increasingly recognise anatomy, monitor performance, display personalised three-dimensional models and automate defined technical tasks.

4. Every clinician may have an AI assistant

A urologist could have a secure system that:

• reviews the patient’s history before clinic;
• summarises previous imaging;
• identifies outstanding investigations;
• checks relevant guidelines;
• drafts the clinic note;
• prepares the correspondence;
• and organises follow-up.

5. Patients may have their own clinical AI

Rather than searching the internet after a consultation, patients might have access to a clinically governed AI assistant connected to their own treatment information and trusted medical sources.

6. Medicine may move from population averages towards individual prediction

This could ultimately be the most important development.

Healthcare has traditionally relied on what happened to groups of similar patients.

AI may allow us to make increasingly accurate predictions about the individual sitting in front of us.

What should patients understand about AI today?

For now, patients should remember several things.

AI is already entering healthcare, but its maturity varies enormously.

A regulated digital pathology tool is not the same as an experimental surgical AI system.

AI does not automatically mean better care.

Technology must demonstrate that it improves outcomes that matter to patients.

The clinician remains responsible.

AI should inform clinical judgement rather than replace it.

Patients should be told when important AI systems are influencing their care.

Transparency will become increasingly important.

And perhaps most importantly:

The objective of medical AI should not be to remove the human relationship from medicine. It should be to give doctors better information and more time to use that relationship effectively.

The future of AI in urology

Artificial intelligence has already progressed beyond being an interesting laboratory technology.

In selected areas of urology, it is moving into clinical practice, particularly in imaging, digital pathology and data-driven prostate-cancer assessment.

Other applications—including AI-assisted cystoscopy, renal-tumour prediction, surgical computer vision and autonomous surgical tasks—remain earlier in development.

Large language models are moving even faster, particularly in documentation and patient communication, sometimes faster than the evidence and governance systems needed to control them.

The next decade is therefore unlikely to involve one dramatic moment when “AI arrives” in urology.

Instead, intelligence will progressively be added to almost every part of the clinical pathway.

The scan will become more intelligent.

The pathology slide will become more intelligent.

The robotic platform will become more intelligent.

The electronic health record will become more intelligent.

And the way we estimate an individual patient’s prognosis may become much more sophisticated.

The challenge for urologists is not to resist that change.

Nor should we accept every new technology simply because it contains artificial intelligence.

Our responsibility is to determine where AI genuinely improves diagnosis, treatment, safety or the patient experience—and where it does not.

The best future is unlikely to be AI replacing the urologist. It is far more likely to be an experienced urologist, supported by increasingly powerful technology, making better-informed decisions with the patient.

---

Need Specialist Advice About a Urological Diagnosis?

Artificial intelligence is becoming increasingly sophisticated, but medical decisions still need to be interpreted in the context of the individual patient.

If you have an elevated PSA, suspected or confirmed prostate cancer, a kidney or bladder abnormality, or another urological problem, I can review your investigations and explain what the findings mean and what should happen next.

As a Consultant Urological Surgeon specialising in prostate cancer, uro-oncology, robotic surgery, and general urology, I see private patients seeking diagnosis, treatment, second opinions, and surgical care in Buckinghamshire, Berkshire, Oxfordshire, and London.

Where appropriate, I can review your:

• PSA history;
• prostate MRI;
• biopsy results;
• CT or other imaging;
• previous treatment recommendations;
• and surgical options.

Arrange a consultation with Mr Neil Haldar for a personalised assessment of your urological condition and treatment options.

Frequently asked questions

---

This article provides general medical information and does not replace individual assessment or advice from an appropriately qualified healthcare professional. The field of artificial intelligence is developing rapidly, and the availability and regulatory status of individual technologies varies between healthcare systems.

Mr Neil Haldar is a Consultant Urological Surgeon specialising in prostate cancer, robotic and open prostate surgery, uro-oncology and general urology. He sees private patients in Buckinghamshire, London and the Thames Valley.