Faculty Bibliography

Deep neural networks (DNNs) show promise in image-based medical diagnosis, but cannot be fully trusted since they can fail for reasons unrelated to underlying pathology. Humans are less likely to make such superficial mistakes, since they use features that are grounded on medical science. It is therefore important to know whether DNNs use different features than humans. Towards this end, we propose a framework for comparing human and machine perception in medical diagnosis. We frame the comparison in terms of perturbation robustness, and mitigate Simpson's paradox by performing a subgroup analysis. The framework is demonstrated with a case study in breast cancer screening, where we separately analyze microcalcifications and soft tissue lesions. While it is inconclusive whether humans and DNNs use different features to detect microcalcifications, we find that for soft tissue lesions, DNNs rely on high frequency components ignored by radiologists. Moreover, these features are located outside of the region of the images found most suspicious by radiologists. This difference between humans and machines was only visible through subgroup analysis, which highlights the importance of incorporating medical domain knowledge into the comparison.

Breast dynamic contrast-enhanced MRI (DCE-MRI) is the most sensitive imaging modality for the detection of breast cancer. Screening MRI is currently performed predominantly in patients at high risk for breast cancer, but it could be of benefit in patients at intermediate risk for breast cancer and patients with dense breasts. Decreasing scan time and image interpretation time could increase cost-effectiveness, making screening MRI accessible to a larger group of patients. Abbreviated breast MRI (Ab-MRI) reduces scan time by decreasing the number of sequences obtained, but as multiple delayed contrast enhanced sequences are not obtained, no kinetic information is available. Ultrafast techniques rapidly acquire multiple sequences during the first minute of gadolinium contrast injection and provide information about both lesion morphology and vascular kinetics. Diffusion-weighted imaging is a noncontrast MRI technique with the potential to detect mammographically occult cancers. This review article aims to discuss the current indications of breast MRI as a screening tool, examine the standard breast DCE-MRI technique, and explore alternate screening MRI protocols, including Ab-MRI, ultrafast MRI, and noncontrast diffusion-weighted MRI, which can decrease scan time and interpretation time.

Pathologic nipple discharge (PND) is typically unilateral, spontaneous, involves a single duct, and is serous or bloody in appearance. In patients with PND, breast MRI can be helpful as an additional diagnostic tool when conventional imaging with mammogram and ultrasound are negative. MRI is able to detect the etiology of nipple discharge in 56-61% of cases when initial imaging with mammogram and ultrasound are negative. Advantages to using MRI in evaluation of PND include good visualization of the retroareolar breast and better evaluation of posterior lesions which may not be well evaluated on mammograms and galactograms. It is also less invasive compared to central duct excision. Papillomas and nipple adenomas are benign breast masses that can cause PND and are well visualized on MRI. Ductal ectasia, and infectious etiologies such as mastitis, abscess, and fistulas are additional benign causes of PND that are well evaluated with MRI. MRI is also excellent for evaluation of malignant causes of PND including Paget's disease, ductal carcinoma in-situ and invasive carcinoma. MRI's high negative predictive value of 87-98.2% is helpful in excluding malignant etiologies of PND.

Internal mammary lymph nodes (IMLNs) account for approximately 10%-40% of the lymphatic drainage of the breast. Internal mammary lymph nodes measuring up to 10 mm are commonly seen on high-risk screening breast MRI examinations in patients without breast cancer and are considered benign if no other suspicious findings are present. Benign IMLNs demonstrate a fatty hilum, lobular or oval shape, and circumscribed margins without evidence of central necrosis, cortical thickening, or loss of fatty hilum. In patients with breast cancer, IMLN involvement can alter clinical stage and treatment planning. The incidence of IMLN metastases detected on US, CT, MRI, and PET-CT ranges from 10%-16%, with MRI and PET-CT demonstrating the highest sensitivities. Although there are no well-defined imaging criteria in the eighth edition of the American Joint Committee on Cancer Staging Manual for Breast Cancer, a long-axis measurement of ≥ 5 mm is suggested as a guideline to differentiate benign versus malignant IMLNs in patients with newly diagnosed breast cancer. Abnormal morphology such as loss of fatty hilum, irregular shape, and rounded appearance (which can be quantified by a short-axis/long-axis length ratio greater than 0.5) also raises suspicion for IMLN metastases. MRI and PET-CT have good sensitivity and specificity for the detection of IMLN metastases, but fluorodeoxyglucose avidity can be seen in both benign conditions and metastatic disease. US is helpful for staging, and US-guided fine-needle aspiration can be performed in cases of suspected IMLN metastasis. Management of suspicious IMLNs identified on imaging is typically with chemotherapy and radiation, as surgical excision does not provide survival benefit and is performed only in rare cases.

Recent data in invertebrates demonstrated that huntingtin (htt) is essential for fast axonal trafficking. Here, we provide direct and functional evidence that htt is involved in fast axonal trafficking in mammals. Moreover, expression of full-length mutant htt (mhtt) impairs vesicular and mitochondrial trafficking in mammalian neurons in vitro and in whole animals in vivo. Particularly, mitochondria become progressively immobilized and stop more frequently in neurons from transgenic animals. These defects occurred early in development prior to the onset of measurable neurological or mitochondrial abnormalities. Consistent with a progressive loss of function, wild-type htt, trafficking motors, and mitochondrial components were selectively sequestered by mhtt in human Huntington's disease-affected brain. Data provide a model for how loss of htt function causes toxicity; mhtt-mediated aggregation sequesters htt and components of trafficking machinery leading to loss of mitochondrial motility and eventual mitochondrial dysfunction.