How Doctors Think (26 page)

My hospital recently created a Web site called "Patient Site" in response to patients' desire for access to their own medical records. All the lab tests and radiology reports are ready for viewing as soon as they are generated. This provides an opportunity for patient and doctor to sit together and go over the results. Reading the language in these reports can be trying—mostly, of course, for patients—so the clinician should point out the radiologist's words that signal a level of uncertainty. The radiologist may have said the area behind the uterus was not well visualized by the scan, or the thickening of the wall of the bowel was not diagnostic for a tumor and could also represent inflammation. This should prompt the clinician to explain to his patient why he needs to revisit the history, perform a more comprehensive physical examination, or order further tests to define the problem.

Communicating this uncertainty poses a challenge. Recently, Orwig viewed a mammogram with a pattern of calcium deposits that are classically diagnosed as benign. But the woman's previous mammogram showed no calcium deposits. He debated with his colleagues about whether to biopsy the breast and came down on the side of biopsy. Orwig went out to speak with the woman whose mammogram showed the new calcium deposits. "I want to apologize in advance," he said. "I think that what we found is benign, but I am going to recommend that it be biopsied." He paused. "I know this will cause you great anxiety. So let me explain my reasoning. What we found on your mammogram reaches the threshold to make this recommendation because it's new. It wasn't present on your prior study. Nonetheless, I am fairly certain that it will turn out to be fine, but we should go ahead, to be complete." It turned out there was a high-grade invasive cancer in the woman's breast, which would be removed by lumpectomy followed by radiation treatment.

Orwig suggested that the case be shown at the quality assurance conferences where the radiologists review their choices and try to refine their skills to avoid future errors. "One colleague said to me: 'If we show this case, given this pattern of calcium deposits, then we are going to have women lining up to the end of the block for biopsies. What good is it going to do? We show this case, people are going to freak out, because then they feel they have to biopsy every patient who has this kind of calcification. We should show cases that will help us because they have very specific findings.'" Orwig agreed with his colleague that, based only on the pattern of calcium deposits, there was no teaching lesson per se, but he believed the key point was changing one's mind based on a previous mammogram. More broadly, he felt that sometimes a teaching point is made by showing the exception to the rule, and emphasizing that there is this gray zone in radiology, particularly in mammography, where judgment comes into play and specific aspects of a case, like the new appearance of calcium deposits, that would otherwise be ignored as benign should be a cause for concern.

Orwig's colleague was afraid that discussing the case would spark an outbreak of "availability errors," the same kinds of errors we saw in the emergency room earlier: a sharp bias in thinking based on a striking, unusual event that recently occurred and becomes prominent in the doctor's mind. Yet, as Orwig argued, not sharing the case could cause colleagues to miss what might be a lethal malignancy. The struggle is to find a middle ground, to be aware of the availability fallacy while recognizing that certain patterns may not conform to the prototype; it is a matter of juggling seemingly contradictory bits of data simultaneously in one's mind and then seeking other information to make a decision, one way or another. This juggling, and this kind of decision-making, marks the expert physician—at the bedside or in a darkened radiology suite.

Orwig thinks about this woman's case often when he is reading mammograms. When he sees a similar pattern of clustered calcium deposits, not only does he check the prior mammogram to see if they were present then, he also checks studies from earlier years to see when the deposits were first noted. Orwig realizes that he could begin to overread mammograms, lowering his threshold to such an extent that he begins recommending unnecessary biopsies. He still is trying to find that middle ground.

Dr. Harold Kundel of the University of Pennsylvania has studied the physiology of image perception by tracking the eye movements of his fellow radiologists. The doctor sits with an apparatus on his head that resembles a bicycle rider's helmet. The apparatus has several parts, including a visor and a miniature video camera. As a doctor examines a series of images, a beam of invisible infrared light is trained on his eye. The camera, trained on his pupil, determines where he is directing his gaze by tracking the infrared beam's reflection. In some of Kundel's studies, a radiologist looked at chest x-rays where there were small lung nodules, each measuring between a half centimeter and one centimeter, or about a quarter to half an inch long. Such nodules are important to detect, since they can represent an early cancer or a serious infection like tuberculosis or fungus. In about 20 percent of the cases, the eyes did not focus at all on the nodules. In the remaining 80 percent, the gaze was directed toward the nodule, but in half of these cases, the nodule simply was not perceived.

"The brain makes a covert decision," Kundel explained. Below the level of consciousness, the mind decides that the image is not important, not worth bringing up to the level of conscious recognition. The radiologists whose eyes dwelled on the nodule for some two to three seconds were more likely to consciously recognize it. Recognition would be enhanced if there was sharp contrast between the nodule and the surrounding lung, the contrast between white for a solid mass and black for air. Recognition was also enhanced if the nodule had a clear border rather than a blurry edge.

Earlier studies tracking eye gaze, done at the University of Iowa, showed that search satisfaction was a common error among radiologists. In follow-up studies, Kundel's team showed that in some cases the gaze did fix on a second abnormality, but it was not recognized. For example, a patient with pneumonia might have a small cancer in the scapula, the wing bone, but the radiologist reported only the pneumonia in the lung, even though the apparatus revealed that his eyes had passed over the tumor in the bone. His mind had already snapped closed after identifying the pneumonia and would not consciously accept other findings. "It comes down to what your preconceived notions of the image are, which I classify as bias," Kundel explained. Echoing a maxim of Merrill Sosman of Brigham Hospital, Kundel said, "You see what you want to see." The expert, though, having learned about bias and search satisfaction, consciously tries to keep his mind open so that he sees beyond his preconceptions. He is helped in this effort by how the clinical history is framed, by the cues provided in the language of the clinician, and by adhering to the kind of systematic deconstruction of the image that Dennis Orwig follows in his dictated reports.

Given the difficulties in perception and cognition that Kundel and other researchers have reported, could computers replace radiologists, or at least lower their error rates? One computer-aided diagnostic system was approved in 2006 by the Food and Drug Administration for identifying lung nodules on chest x-rays. Other systems are being studied, including those for mammograms. The pivotal clinical trial on malignant lung nodules that led to the FDA approval involved fifteen radiologists who were asked to note their level of suspicion that a chest x-ray contained cancer. They used a scoring system of 1 to 100, and they were to mark the location that caused their suspicion. Eighty cancer cases and 160 cancer-free cases were in the study. Each radiologist interpreted these 240 cases three times: two times separated by one to four months without computer assistance and then, immediately after the second interpretation, with computer assistance. Computer assistance improved detection of the cancer between 14 and 24 percent, depending on its size. But the computer system also caused radiologists to change almost 10 percent of their correct decisions (identifying the cancer) to incorrect diagnoses (stating that it was unimportant or benign). Of the fifteen radiologists in the clinical trial, no two had identical results in evaluating the 80 cancer cases and 160 cancer-free cases. All but 25 percent of the cancers were identified by all fifteen radiologists. But the difficult-to-diagnose cancers were found by only four of the radiologists. No radiologist identified all 80 of the cancers correctly.

One unwelcome effect of computer-assisted detection was that after being prompted by the computer, more radiologists suspected cancer in chest x-rays that came from patients without a malignancy—a false-positive reading. This demonstrates the power of technology, particularly computer-based, in shaking the confidence of a specialist in his initial diagnosis. It also demonstrates that machines do not provide perfect solutions to the imperfection of perception and thinking. Perhaps, as radiologists become more accustomed to computer-assisted detection and receive clinical feedback about the risk of becoming overly suspicious about benign findings on a chest x-ray, they will accommodate their thinking to the new technology. In the meantime, as they search for another new middle ground, there will be a tradeoff, with more accurate cancer detection but greater patient anxiety, as more people without cancer are subjected to the emotional upheaval and invasive procedures that follow on false positives.

A short drive south from Marin County, across the Golden Gate Bridge, brings you to San Francisco. Perched on Parnassus Heights is the University of California Medical Center and the nearby Moffitt Hospital. Vickie Feldstein is a professor of radiology at UCSF specializing in ultrasonography. (Dr. Dennis Orwig happens to be her husband.) Most people are familiar with ultrasound examinations from a pregnancy. The developing baby is imaged, appearing in a two-dimensional representation inside the uterus, a swirl of black and white and gray. "Some people consider an ultrasound image to look like a weather map," Feldstein said with a chuckle. It certainly looks like a weather map to me, specifically a snowstorm. The flux of white specks across a black background makes the discrete outlines of organs difficult, if not impossible, for me to make out. Of course, for Feldstein and radiologists who use this technology daily, the images are as familiar as the palms of their hands, and the contrasts of black, white, and gray full of meaning.

Given the complexity of the images in ultrasonography, one might think that in this case computers would better assist diagnosis. The computer would provide a quantitative assessment of each structure in the developing fetus. For example, at twenty weeks' gestation, an ultrasound is used to measure the ventricles of the fetus, which are the fluid-filled cisterns in the brain. If the length of the ventricle is greater than ten millimeters, then the fetus is carefully monitored for hydrocephaly, commonly called water on the brain, a disorder of ballooning ventricles that can result in brain damage as well as other developmental abnormalities. But it turns out that the numbers a computer would supply may not reveal what the radiologist wishes to consider. "The numbers will help you and raise your level of attention," Feldstein said, "but you have to take the whole picture. You have to look at the shape of the ventricles and the associated surrounding tissue. It's not just based on reading the numbers."

A normal cerebral ventricle is shaped like a teardrop. The ventricle on an ultrasound is defined by a black central core, which is the fluid, and a white lining, the choroid plexus that produces the fluid. Feldstein recently saw a woman close to her due date. "She was near term," she recalled, "and the size of the fetus's ventricles was within the numerical limits of normal. But when I looked at it, the shape didn't seem right." The changes in the contour of the teardrop were subtle, but to Feldstein's trained eye, potentially significant: the borders were not smooth but slightly irregular, and the teardrop was not finely tapered at its apex. Both of these observations could be easily ignored or discounted, particularly since the dimensions were not beyond the accepted limits of health. Feldstein decided that she needed to pursue this, although the clinical consequences of her findings were not immediately clear. At thirty-five weeks, the question in her mind was, What would the mother do with the information? It was too late to consider terminating the pregnancy, but Feldstein concluded that it was important to know whether her impression—that there was some underlying abnormality in the baby's brain—was correct or not. In part, determining this would help the parents anticipate problems after birth, preparing them emotionally and logistically for raising a child that might be retarded or need special neonatal care.

Another dimension that influenced Feldstein's decision-making: the medical-legal ramifications. If indeed there was an abnormality in the brain that was not visualized on the ultrasound but had caused the fetus's ventricles to change their contour, then it would be best to know that before delivery so no one would suggest that an inept obstetrician had caused trauma that led to brain damage. Feldstein explained to the mother that although her baby's ventricles were within the limits of normal, there also were subtle changes in shape that might bespeak something abnormal in the brain. Feldstein didn't want to unduly frighten the mother, but on the other hand, she felt it was her responsibility to communicate her analysis. The mother decided to undergo MRI scanning, and a cerebral hemorrhage was detected. Bleeding in the fetus's brain had caused the ragged borders of the ventricles and the distortion of the apex of the teardrop. Feldstein's sharp eye had been proven true. The mother delivered her baby with the necessary pediatric neurologists in attendance.

Every radiologist I spoke to could immediately recount not only successes like Feldstein's but also unnerving errors. Herbert Kressel, the imaging specialist at Harvard, told me that recently he had missed seeing an abnormality on an MRI scan: a small but discernible cancer of the liver that was present on several of the images. "It was a definite miss. I just didn't see it. And to this day, I really don't know why." He wondered whether he had moved too quickly through the cine presentation, put too much pressure on the tracker ball. "But that's a speculation. I just don't know," he said, his voice heavy. "People have to understand that there always will be a certain amount of imprecision in imaging and interpretation."