---

Could AI Help Personalise Prostate Cancer Treatment?

By Mr Neil Haldar, Consultant Urological Surgeon

Last medically reviewed: July 2026

In brief

Artificial intelligence could help personalise prostate cancer treatment by combining information from biopsy images with clinical data such as PSA, tumour stage and age.

Some AI models can estimate the risk of cancer spreading more accurately than conventional risk groups. More importantly, early research suggests that certain models may help predict whether an individual patient is likely to benefit from adding hormone therapy to radiotherapy.

This is an important distinction. A test that predicts how aggressive a cancer may be does not necessarily predict whether a particular treatment will help.

AI-based biomarkers are promising, but they do not currently replace an experienced multidisciplinary team. They must be considered alongside MRI and scan findings, biopsy results, the patient’s general health, treatment side effects and personal priorities.

Why does prostate cancer treatment need to be personalised?

Prostate cancer is not one single disease.

Some prostate cancers grow so slowly that they may never cause harm. Others are more aggressive and may spread without effective treatment. Even cancers with apparently similar PSA levels, stages and biopsy grades can behave differently.

This creates several difficult decisions:

• Is active surveillance safe, or is treatment advisable?
• Should treatment involve surgery or radiotherapy?
• Should hormone therapy be added to radiotherapy?
• How long should hormone therapy continue?
• Is additional treatment needed after surgery?
• How intensively should a patient be monitored?
• Could additional treatment cause more harm than benefit?

Current UK guidance divides localised prostate cancer into clinical prognostic groups using factors such as PSA, clinical stage and Grade Group. Treatment choices can include active surveillance, radical prostatectomy and radical radiotherapy, depending on the risk category and the individual patient’s circumstances.

These classifications are useful, but no two men in the same group are biologically identical.

How are prostate cancer treatment decisions made now?

Doctors currently combine several sources of information.

These may include:

• PSA level and PSA changes over time;
• MRI findings;
• prostate-biopsy Grade Group;
• the amount of cancer present in the biopsy;
• clinical or pathological stage;
• PSMA PET or other staging scans when appropriate;
• age, health and life expectancy;
• urinary and sexual function;
• family history and inherited genetic changes;
• the patient’s attitude towards risk and treatment side effects.

The case may then be reviewed at a multidisciplinary team meeting involving urologists, radiologists, pathologists, oncologists and specialist nurses.

This is already a form of personalised medicine. The limitation is that people can only interpret a finite number of variables, while some important information may be hidden within the microscopic appearance of the cancer.

AI may detect and combine patterns that are difficult for the human eye to quantify.

How could AI personalise prostate cancer treatment?

One of the most promising approaches is known as multimodal artificial intelligence, or MMAI.

A multimodal model analyses more than one type of information. In prostate cancer, this might include:

• a digitised image of the prostate biopsy;
• PSA;
• age;
• tumour stage;
• Grade Group;
• previous treatments;
• long-term outcomes from similar patients.

The AI examines thousands or millions of visual features across a digital biopsy slide. These may include patterns in the shape, arrangement, and surrounding environment of cancer cells that pathologists do not routinely measure.

It then combines these image features with the clinical information to calculate an individual score.

The aim is not simply to label the cancer as low, intermediate or high risk. The longer-term ambition is to answer more clinically useful questions, such as:

How likely is this cancer to spread?

Is this patient likely to benefit from hormone therapy?

Does the probable benefit justify the potential side effects?

What is digital pathology?

A pathologist examines a traditional prostate biopsy under a microscope.

With digital pathology, the glass slide is scanned at very high resolution. The pathologist can review the tissue on a computer screen, while AI software can analyse the same image.

An AI system may identify patterns involving:

• cancer-cell shape;
• gland structure;
• tumour architecture;
• the relationship between cancer and surrounding tissue;
• variation within different areas of the tumour.

The system does not necessarily need to discover a completely new biological marker. It can learn how combinations of visual features relate to outcomes observed during long-term clinical trials.

The final pathology diagnosis and Grade Group should still be determined by a qualified pathologist.

What is the difference between a prognostic and predictive AI test?

This distinction is essential.

A prognostic test

A prognostic test estimates what is likely to happen.

For example:

This cancer has a relatively high risk of recurring or spreading over the next ten years.

That information may support closer monitoring or more intensive treatment, but it does not prove that one particular treatment will provide greater benefit.

A predictive test

A predictive test estimates whether a specific treatment is likely to change the outcome.

For example:

Patients with this biomarker result appear to gain a meaningful reduction in the risk of distant spread when hormone therapy is added to radiotherapy.

A truly predictive biomarker may be more valuable because it can help doctors intensify treatment for people likely to benefit while sparing others unnecessary side effects.

Can AI predict who will benefit from hormone therapy?

This is currently one of the strongest examples of AI-based prostate cancer treatment personalisation.

Androgen deprivation therapy, commonly called hormone therapy, is often combined with prostate radiotherapy. It can improve cancer control in appropriately selected patients, but it may also cause side effects including hot flushes, fatigue, loss of sexual function, changes in body composition, reduced bone density and metabolic effects.

Researchers developed an AI model using digitised pretreatment biopsy images and clinical information from 5,727 patients enrolled in five randomised phase III trials.

The locked model was then tested using 1,594 patients from a separate randomised trial comparing radiotherapy alone with radiotherapy plus four months of hormone therapy.

Approximately one-third of patients were classified as biomarker-positive. Within this group, adding hormone therapy was associated with a substantial reduction in the risk of distant metastasis. Among biomarker-negative patients, the study found no statistically significant difference between the two treatment groups. The biomarker-treatment interaction was statistically significant.

This is important because the model was validated using data from a randomised treatment trial rather than relying solely on retrospective hospital records.

However, the findings should not be interpreted as meaning that every biomarker-negative patient can safely avoid hormone therapy. Treatment decisions must consider the full clinical picture, and prospective studies are examining how the test affects real-world shared decision-making.

Could AI help decide how long hormone therapy should continue?

Possibly.

For men with higher-risk prostate cancer receiving radiotherapy, the duration of hormone treatment can be a difficult balance. Longer treatment may improve cancer control for some patients but also prolongs its physical, sexual and metabolic side effects.

A 2025 study developed an AI digital-pathology biomarker using biopsy images and clinical data from six randomised phase III trials. It was independently validated in a seventh randomised trial that compared four months with 28 months of hormone therapy.

The purpose was to identify patients with high-risk prostate cancer who were more likely to benefit from the longer course.

This represents an important direction for personalised treatment. Rather than deciding treatment duration using a broad risk category alone, future decisions may also be informed by the biological patterns within an individual patient’s biopsy.

Further prospective implementation evidence will still be needed before such testing can be regarded as routine across all healthcare systems.

Could AI help choose between active surveillance and treatment?

Potentially, but the evidence is less direct.

Active surveillance aims to avoid or delay surgery or radiotherapy in men whose cancer appears unlikely to cause harm. It involves scheduled PSA tests, clinical review, MRI and sometimes repeat biopsy.

The difficulty is identifying:

• cancers that are genuinely suitable for surveillance;
• cancers that have been underestimated by the biopsy;
• patients whose disease is more likely to progress;
• the safest intensity of follow-up.

A multimodal AI biomarker originally developed using North American radiotherapy trial data was subsequently evaluated in a prospective Swedish cohort of 143 men who underwent radical prostatectomy.

The score was associated with adverse findings at surgery and with later biochemical recurrence. Estimated five-year recurrence rates were 25% in the intermediate-to-high AI groups compared with 4% in the low group.

This suggests that AI may identify additional information in the original biopsy. However, this study did not test whether using the result to select or exclude men from active surveillance improves patient outcomes.

AI should therefore not currently be used by itself to decide that treatment is necessary.

Can AI decide whether surgery or radiotherapy is better?

Not at present.

An AI system may estimate cancer aggressiveness or the likely benefit of adding a particular treatment. That is different from proving whether radical prostatectomy or radiotherapy is the better choice for an individual patient.

Surgery and radiotherapy have different practical implications and side-effect patterns. The most appropriate option depends on factors including:

• cancer stage and location;
• prostate size;
• urinary symptoms;
• bowel health;
• existing erectile function;
• previous abdominal or pelvic treatment;
• age and general health;
• the patient’s personal preferences.

An algorithm cannot fully capture how an individual patient values urinary control, sexual function, avoiding medication, treatment convenience, pathological certainty or the possibility of future salvage treatment.

AI may eventually improve comparisons between treatment strategies, but it should support, not replace, shared decision-making.

Could AI help after radical prostatectomy?

After prostatectomy, PSA should usually fall to an undetectable or very low level. A subsequent rise may indicate biochemical recurrence.

The next decision can be difficult. Some patients may benefit from early salvage radiotherapy, while hormone therapy may be added for selected higher-risk cases.

Researchers have developed an MMAI model using prostatectomy tissue images and clinical variables from men enrolled in the NRG/RTOG 9601 and 0534 randomised trials.

In its validation cohort of 533 patients, the AI score was independently associated with the risk of distant metastasis. Ten-year metastasis incidence was approximately 25% in the AI-defined high-risk group and 8.8% in the low-risk group.

The apparent absolute benefit from adding hormone treatment was also larger in the high-risk group, although the authors concluded that further external and prospective validation was still underway.

This is a promising application, but it should not yet be interpreted as definitive proof that hormone therapy can safely be omitted solely on the basis of the AI result.

Could AI combine MRI, pathology and genetic information?

This is likely to be the next major stage.

A future prostate cancer model could potentially combine:

• multiparametric MRI;
• PSMA PET;
• digitised biopsy or prostatectomy pathology;
• PSA and PSA kinetics;
• inherited genetic variants;
• molecular changes within the cancer;
• response to previous treatment;
• long-term outcomes from large patient populations.

Current multimodal models have mainly combined pathology images with a relatively small number of clinical variables. Combining imaging, pathology and molecular data may produce a more complete biological description of the cancer.

However, adding more information does not automatically produce a better clinical test. The model must still be externally validated, calibrated for the population in which it is used and shown to improve real treatment decisions.

Is AI-based personalised treatment available in the UK?

Availability is limited and varies between healthcare systems and centres.

AI-derived prostate cancer biomarkers have been developed commercially and have entered clinical use in parts of the United States. Prospective implementation studies are assessing how one such test affects conversations about hormone therapy with radiotherapy.

These tests are not currently a routine component of the standard NICE prostate cancer pathway. Current UK guidance continues to base treatment recommendations on recognised clinical prognostic groups, imaging, pathology, patient health and shared decision-making.

A test being commercially available or included in an overseas guideline does not necessarily mean that it is funded, routinely accessible or clinically appropriate for every UK patient.

What are the potential benefits?

Used appropriately, AI could help:

• distinguish cancers needing treatment from those suitable for surveillance;
• provide more individual estimates of recurrence and metastasis risk;
• identify patients more likely to benefit from hormone therapy;
• reduce avoidable treatment and side effects;
• guide the intensity and duration of treatment;
• improve conversations between patients and their clinical teams;
• make advanced pathology analysis more scalable.

The most valuable outcome would not be a more complicated risk score. It would be a clearer treatment decision that improves cancer control or quality of life.

What are the limitations and risks?

AI may be accurate but not clinically useful

A model can predict recurrence more accurately without changing which treatment should be recommended.

Research populations may not represent every patient

A model developed from historic radiotherapy trials may perform differently in contemporary UK patients, in different ethnic groups or after other forms of treatment.

An analysis across five randomised trials found that two MMAI prognostic models retained performance in African American and non-African American groups, but fairness must be assessed continually rather than assumed.

Pathology processing can differ

Slide preparation, staining and scanning equipment vary between laboratories. These differences may affect algorithm performance.

Treatments change over time

MRI, PSMA PET, radiotherapy techniques, surgery and systemic treatments continue to develop. A model trained on older trials must be shown to remain relevant to current practice.

A probability is not a certainty

A low-risk score does not guarantee that cancer will not progress. A high-risk score does not mean that progression is inevitable.

Commercial conflicts must be transparent

Many important AI biomarker studies involve researchers employed by or connected with the companies developing the tests. This does not invalidate the findings, but independent evaluation and transparent disclosure are important.

Personal preferences cannot be reduced to an algorithm

A treatment that provides a small statistical benefit may be worthwhile to one patient and unacceptable to another.

What should a patient ask about an AI prostate cancer test?

Useful questions include:

1. What exact decision is the test intended to help with?
2. Is it prognostic, predictive or both?
3. Has it been validated in a separate group of patients?
4. Was it tested using a randomised treatment trial?
5. Does the evidence apply to my stage and proposed treatment?
6. Could the result change my treatment recommendation?
7. What are the consequences if the result is wrong?
8. Is the test recommended or funded in the UK?
9. Who developed the test, and were the studies independently evaluated?
10. Will my case still be reviewed by a multidisciplinary team?

Will AI replace the prostate cancer multidisciplinary team?

No.

A sophisticated AI model may recognise patterns across thousands of biopsy images and clinical records. It cannot independently examine a patient, assess surgical feasibility, understand family circumstances or take responsibility for the consequences of treatment.

The safest model is:

AI evidence + specialist interpretation + multidisciplinary review + informed patient choice

The doctor’s role may change as better computational tools become available, but clinical responsibility should remain clearly human.

Conclusion

Artificial intelligence has genuine potential to make prostate cancer treatment more personal.

The strongest evidence currently concerns AI models that combine digital biopsy images with clinical information. These models can improve risk assessment and, in selected radiotherapy settings, may help identify which patients are more likely to benefit from hormone therapy.

Further research is exploring whether AI can guide hormone-treatment duration, improve active-surveillance selection and personalise treatment after a rising PSA following prostatectomy.

However, AI cannot yet reliably decide between surgery and radiotherapy, guarantee an individual outcome or replace multidisciplinary judgement.

The appropriate question is therefore not:

Can AI choose my prostate cancer treatment?

It is:

Can carefully validated AI provide additional information that helps my clinical team and me make a better-informed decision?

For selected patients, the answer is increasingly likely to be yes.

---

Making a Prostate Cancer Treatment Decision?

No AI test, risk score or algorithm should be considered in isolation. The most appropriate treatment depends on the cancer’s biology and extent, your general health, urinary and sexual function, and what matters most to you.

If you have recently been diagnosed with prostate cancer, have been offered surgery or radiotherapy, or would like a second opinion before deciding on treatment, Mr Neil Haldar can review your results and discuss the options with you in detail.

This may include consideration of your PSA, MRI, biopsy findings, staging investigations and the potential benefits and side effects of treatments including active surveillance, radical prostatectomy and radiotherapy.

Frequently asked questions

---

About the author

Mr Neil Haldar is a consultant urological surgeon with a specialist interest in prostate cancer, radical prostatectomy, robotic surgery and emerging technologies in urological care.

---

Medical disclaimer

This article provides general information and does not replace individual medical advice, examination or multidisciplinary assessment. Prostate cancer treatment should be discussed with an appropriately qualified urologist or oncology team familiar with the patient’s complete clinical circumstances.