Phase II trials study the safety and effectiveness of the drug in cases with the indication in question, and seek to identify the optimum dosage schedule. Phase II studies, which typically involve 100 to 200 subjects, can also investigate the pharmacological differences between patients and healthy volunteers.

The phase III study is the most extensive phase of development. Its purpose is to confirm the effect, tolerability and safety of the new drug compared with standard therapy (or placebo), and to prove the correct dose in cases with the indication in question. Phase III studies usually involve 500 to 5000 patients, and patient demographics should as far as possible be representative of the population in which the new drug will ultimately be used. Phase IIIb studies are not needed for registration, but typically answer specific questions in order to support the new drug’s use in particular populations or territories.

Phase IV studies are needed after registration. Many questions may remain unanswered after the phase III study, such as effectiveness and safety in children or the elderly. Large post-marketing surveillance studies are now often conducted on new drugs to identify rare adverse events.

Clinical trials have historically used clinical endpoints (outcomes) to establish whether the therapy is safe and effective. A clinical endpoint is defined as a characteristic or variable that measures how a patient feels, functions or survives. The classic endpoint of mortality determines whether the new therapy decreases the rate of death in comparison with that of a control group. Another endpoint is morbidity, in which investigators examine whether patients undergoing the therapy are in some way more functional or enjoy a higher quality of life than those who do not receive the therapy. Trials with these clinical outcomes often have a long duration and require a large number of subjects, and are therefore extremely costly.

A surrogate endpoint is a laboratory measurement or physical sign used as a substitute for a clinical endpoint (4,5). Regulatory approval for a drug may be secured when the regulatory authority decides that the product has an effect on a surrogate endpoint that is reasonably likely to predict clinical benefit (6). It has been shown that small improvements in clinical trial outcomes and decision-making translate into great cost-savings and a faster time-to-market (7). Examples of surrogate endpoints include: (i) blood pressure as a risk for myocardial infarction or stroke; (ii) cholesterol measurements for risk of myocardial infarction or death; (iii) CD4 cell count for risk of progression to AIDS; and (iv) CT or MRI measures of tumour size. In a survey of oncology drug approvals by the FDA, of the 57 oncology drugs approved for marketing during the period of 1990-2002, only 18 relied upon traditional survival data. Of the additional 14 drugs that gained accelerated approvals, none relied upon traditional survival data. The clinical trials of 53 out of 71 oncology drugs used surrogate endpoints, the most common one being the change in tumor size as a result of the drug therapy (3). On many occasions, the use of a surrogate endpoint shortens substantially the total time required for confirmation of clinical benefits.

True surrogate endpoints will require a comprehensive validation, i.e. strong evidence of positive predictive clinical outcome. Approval under this section may be subject to the requirement that the sponsor study the drug further, to verify its clinical benefit. However, in some instances surrogate endpoints used to assess the efficacy of new drugs were found not to predict the clinical outcomes of mortality or morbidity, leading to the withdrawal of those drugs. Among various explanations for such a failure is that the disease process could affect the clinical outcome through several causal pathways that are not mediated through the surrogate. The drug might also affect the clinical outcome by unintended and unrecognized mechanisms of action that operate independently of the disease process and are not recognized by the surrogate endpoint (8). One notable example is the reliance on suppressing ventricular dysrhythmia determined from electrocardiograms for the development and approval of anti-arrhythmia drugs, the result of which is the marketing of drugs that were later found to cause mortality at two to two and a half times higher than that of the placebo group (9,10).

PET has been applied to a wide number of drugs to demonstrate activity in vivo, from standard chemotherapy such as 5-fluorouracil to molecular agents such as those involved in tumour angiogenesis and antivascular therapy (13). The use of PET imaging techniques to establish dosing regimens has been pursued (14,15). PET can be applied before traditional Phase 1 studies to test compounds in humans at tracer (non-pharmacologically active) concentrations. Such an approach uses as little as one-thousandth of the starting dose (i.e. micro-dosing) of a typical Phase 1 study. In broad terms, imaging of PK properties falls into two categories. The first category involves the radiolabeling of compounds that interact with, or neutralize, agents from the environment, such as toxins, bacteria and viruses. In this case, generally only tissue concentrations of drugs are necessary. In the second category, if the drug is expected to alter or modulate some aspects of the pathophysiologic process, then imaging studies are generally used to characterize the number of receptors, binding efficiency and receptor occupancy (11). As an example of the first category, the development of 18F-labeled antifungal agent fluconazole (Diflucan, Pfizer) was monitored by PET to establish the concentration of the drug in different organs, particularly at the site of infection. The imaging study found that the observed concentrations compared favorably to the concentrations required to inhibit in vitro pathogen growth (11, 16). PET imaging of aprepitant (Emend, Merck) belongs to PK imaging of the second category. Aprepitant is a neurokinin-1 (NK-1) receptor antagonist that crosses the blood-brain barrier and was developed as, among other things, a treatment for emetogenic chemotherapy-induced nausea and vomiting. By using a 18F-labeled ligand with known high affinity and specificity for the NK-1 receptor, PET was used to image the displacement of this radioligand by aprepitant. During the clinical trial, because NK-1 receptors have been found to be most abundant in the caudate and putamen, and least abundant in the cerebellum, this information was used to establish a reference and to compute the displacement of the PET tracer (11,17).

Imaging modalities other than PET have been used to evaluate dosing regimens, for example, MRI was used to evaluate drug regimens for interferon-β (IFN-β) treatments of multiple sclerosis (MS) (18), infliximab (Remicade, Centocor) for psoriatic arthritis (19), and PTK787 for colorectal cancer (20).

Etanercept (Enbrel, Immunex, now Amgen) is a tumor necrosis factor inhibitor for the treatment of rheumatoid arthritis. When examining the potential for etanercept as a first-line treatment, early trials used two sets of criteria: (i) American College of Rheumatology (ACR) scores, which use a combination of subjective pain and function assessments, in addition to serum C-reactive protein levels; and (ii) conventional radiography images of joint-space narrowing and erosion. Whereas clinical scoring showed no significant difference between etanercept and methotrexate (the standard therapy at the time), the imaging-based erosion score showed statistically significant differences (28). On the basis of these data, the FDA granted Immunex marketing approval with the condition that additional supporting data be collected. A subsequent study showed etanercept achieved sustained improvements over methotrexate in terms of both clinical and imaging scores (29).

The current gold standard for assessing outcomes of Alzheimer’s disease comprises behavioural or cognitive measures, but these suffer from poor reliability. MRI measurement of whole brain or hippocampal atrophy rate can be used to support clinical outcome measures in therapeutic trials for Alzheimer’s disease, and functional brain activity can be objectively quantified by measuring regional glucose metabolism with PET (30,31,32). With clinical trials for MS, there has been great reliance on imaging data, especially using T2-MRI lesion burdens, and the number of contrast-enhancing lesions. These imaging biomarkers can serve as primary outcome measure in Phase I and Phase II trials, and also can serve as secondary outcome in Phase III trials (33,34,35). Other examples of early bioactivity assessment include MRI and CT for examining ischemic stroke (36,37) and imaging assessment of cardiovascular disease (38).

Imaging can also help to detect early disease and define stratified study groups. Imaging can be used to separate - as early as possible - responders from non-responders in patients undergoing therapeutic intervention. Many diseases have a “therapeutic window” in their course, during which medical intervention may have a more significant impact (39). In clinical trials, excluding patients who are not likely to progress can substantially increase the statistical power of a study and thereby reduce the number of patients and study duration needed to prove therapeutic efficacy. In one study examining advanced non-small-cell lung cancer, 55 patients were imaged with FDG-PET after a single course of chemotherapy. The results showed a statistically significant difference both in time to progression and overall survival between responders (i.e. those with observed decrease in tumor metabolism as seen on PET) and non-responders (40). Similarly, in a study involving 40 gastroesophageal cancer patients, FDG-PET segregated responders from non-responders with a sensitivity of 93% and specificity of 95% (41).

The heart has limited capacity to repair itself. Toxic findings in the heart can be serious. While electrical activities of the heart can be monitored in clinical trials by ECG, due to delayed release of serum markers of cardiac damage, structural histopathology, such as cardiomyocyte inflammation, degeneration and necrosis lack conventional early biomarkers. Echocardiography has been widely used in preclinicaland clinical drug safety evaluation for new drug cardiotoxicity (45,46). Echocardiography can be used to measure the cardiac wall thickness, lumen volume and cardiac output. It has further advantages that it provides low cost, real-time images. MRI has also been used to determine myocardial volume, wall thickness, left and right ventricular end-diastolic and end-systolic lumen volumes, stroke volume and ejection fraction.

Many non-invasive tests of kidney function can only show renal damage after the functional reserve had been eliminated. This reserve can compensate for up to 75% of the loss, which make many serum and urine biomarkers insensitive for early kidney damages. MRI can offer advantages over methods that measure global functional changes by providing anatomically specific information of kidney injury. For example, a wide range of compounds can cause renal papillary necrosis (RPN). In the early development of RPN, there are few clinical symptoms and specific urine or blood biomarkers. The progression of renal damage is insidious and renal function may be severely compromised before the condition becomes obvious. The diagnosis of RPN tends to be made in the late stages of this disease after irreversible destructive changes have occurred. The use of imaging modalities has led to an increased positive diagnosis in human population (47). Lang et al reported that contrast-enhanced multi-phasic CT scan can identify early manifestations of RPN and medullary necrosis, and CT scan can further be used to monitor lesion progression or regression after treatment (48).

The absence of a reliable clinical safety biomarker can lead to the withdrawal of an attractive new drug from further clinical study. Vigabatrin (Sabril), an irreversible inhibitor of gamma-aminobutyric acid transaminase, is an effective treatment for refractory epilepsies. Animal toxicology showed that administration of Vigabatrin induces intra-myelinic oedema and microvacuolation in discrete brain regions in rats and in dogs which are detectable with MRI (49,50). Peyster et al reported when Vigabatrin was withdrawn, both MRI abnormality and pathological changes began to decrease, and 16 weeks after vigabatrin withdrawal, MRI and histopathology returned to normal (50). Therefore, MRI being a safety biomarker for the surveillance of Vigabatrin -induced neuropathy enabled further clinical trials of Vigabatrin. Throughout development and post-marketing phase, MRI and neuropathological studies of patients exposed to long-term Vigabatrin treatment have provided no evidence of neuropathy (51). Later, it was concluded that the neurotoxicity of vigabatrin bears a species specificity, with rats and mice being highly susceptible, dogs moderately susceptible and primates and humans not significantly affected at all.

The main aim of image processing in clinical trials is to quantitatively extract parameters objectively by segmentation of tissue structures and derivation of parameters from the selected region of interest. The need for robust objective image analysis is demonstrated in the study of tumour size by Erasmus et al (59). Intra- and inter-reader performance was assessed using CT images from a study of non-small-cell lung cancer. The inter-observer variability in maximum tumour diameter was < 7.1% for a well-marginated mass vs < 140% for a poorly marginated mass. Therefore tumour response could be misinterpreted owing to the inconsistent definition of lesion size, particularly for inter-reader variability. If such variability occurred within a clinical trial, then it may not be possible to infer whether the drug worked or not (3).

Blinded image evaluation is usually the favored approach. Sometimes unblinded image evaluation is used to show consistency with the results of fully blinded image evaluation. Two blinded readers (or preferably three or more) should evaluate images. To prevent bias, it is preferable to involve a number of readers from differing centres, and readers may need to be excluded from reading cases from their own institutions. Ideally, each reader should view all of the images. In large studies, where it may be impractical to have every image read by each reader, a chosen subset of images may be selected for the assessment of interobserver agreement. Consistency among readers should be measured quantitatively (eg with the kappa statistic). Reader disagreement and inter- or intraobserver variation can often be minimized by a training period. The image readers’ performance should be monitored and documented. In evaluating images, objective, quantifiable endpoints should be used whenever possible. This can result in a higher inter- and intraobserver correlation than qualitative evaluation (60). Items on the image evaluation case report forms should be carefully constructed to gather information without introducing a bias that indicates the answer that is being sought.

In clinical trials with imaging components, a subject is typically scanned at different time points. Image alignment using registration algorithms simplifies the interpretation and correlation of findings between such studies by removing the effect due to difference in patient placement. During image assessment, images acquired at different times should be displayed using a standard format. It has been shown that the sequence in which the images from different time points in the study are viewed may have an impact on the sensitivity and reliability of the evaluation (61). Offsite image evaluations are performed at sites that have not otherwise been involved in the conduct of the study, and by readers who have not had contact with patients or other individuals involved in the study (62). Centralized offsite reading can support more complex scoring methods and quantitative analyses than are feasible in clinical practice. Trials intended to demonstrate or support efficacy generally should use offsite image evaluations at a limited number of sites (or preferably at a centralized site).

Computer-aided detection (CAD) can improve readers’ performance in detection of abnormalities, as well as in characterization of detected abnormalities. Manual delineation of a structure, particularly on a 3D image series, is slow and expensive. Thus a variety of computer-assisted methods is used to identify structures of interest, and to extract information such as lesion number, area, volume, density and intensity. These computer systems need to find an appropriate balance between the amount of user interaction and automation, which improves consistency, reduces the time to completion and enables clinical experts to bring their expertise into the analysis procedure (63). Potential time saving and reduction in measurement variability achieved by automated algorithms should not be allowed to compromise the accuracy of the analysis (64).