BRAF V600E mutation has emerged as a marker of aggressive behavior in papillary thyroid carcinoma but its significance in microcarcinoma is not entirely clear. One-hundred and twenty-nine papillary thyroid microcarcinomas were tested for BRAF V600E mutation by single-strand conformation polymorphism, and their clinicopathologic features (age, sex, tumor size, multifocality, nodal metastases, histologic subtype, tumor cell morphology, architecture, tumor-associated stromal reaction, tumor interface to non-neoplastic thyroid (well circumscribed vs infiltrative), extrathyroidal extension, lymphovascular invasion, intratumoral multinucleated giant cells, and adjacent non-neoplastic thyroid pathology) were examined. Compared with tumors without the mutation (39/129, 30%), the mutated microcarcinomas (90/129, 70%) showed significantly higher prevalence of infiltrative tumor borders (78/90 vs 23/39, P ¼ 0.001), tumor-associated stromal desmoplasia/fibrosis and/or sclerosis (80/90 vs 25/39, P ¼ 0.002), classic nuclear features of papillary thyroid carcinoma (90/90 vs 35/39, P ¼ 0.008) and cystic change (43/90 vs 11/39, P ¼ 0.05). BRAF V600E mutation was more frequent in classic (75%), tall cell (91%), and other variants (470%) than in follicular variant (21%) of papillary thyroid microcarcinoma. Tumors without the mutation were significantly more likely to be solid, well circumscribed, and lacked desmoplasia/fibrosis or sclerosis. However, on multivariate analysis, only the follicular variant of papillary microcarcinoma was significantly associated with the absence of mutation (odds ratio (95% confidence interval): 0.09 (0.01-0.54)). Lymph node metastases (n ¼ 24) were more frequent in microcarcinomas with mutation than without (21/24 vs 3/24, P ¼ 0.02). All patients with lateral cervical node metastasis (n ¼ 9), and all but one tumor with extrathyroidal extension (n ¼ 17/18) showed BRAF V600E mutation. No significant differences were noted in age, sex, tumor size, multifocality, lymphovascular invasion, psammoma bodies, stromal calcification, intratumoral multinucleated osteoclastic-type giant cells, and lymphocytic infiltration between the two groups of tumors. BRAF V600E mutation is an early event in thyroid carcinogenesis, and is associated with distinctive morphology and aggressive features even in papillary thyroid microcarcinomas.

BRAF V600E mutation-positive and negative PTCs are morphologically different. Recognition of their morphology may help in the selection of appropriate tumours for genetic testing.

In a small subset of mucinous carcinomas (20% in this series), positive labeling for CK5/6 indicated primary cutaneous tumor. Staining with p63 also favored primary over metastatic disease. Myoepithelial cell layers were not consistently identified to enable the identification of primary disease.

Low-precision representation of deep neural networks (DNNs) is critical for efficient deployment of deep learning application on embedded platforms, however, converting the network to low precision degrades its performance. Crucially, networks that are designed for embedded applications usually suffer from increased degradation since they have less redundancy. This is most evident for the ubiquitous MobileNet architecture [10,20] which requires a costly quantization-aware training cycle to achieve acceptable performance when quantized to 8-bits. In this paper, we trace the source of the degradation in MobileNets to a shift in the mean activation value. This shift is caused by an inherent bias in the quantization process which builds up across layers, shifting all network statistics away from the learned distribution. We show that this phenomenon happens in other architectures as well. We propose a simple remedy -compensating for the quantization induced shift by adding a constant to the additive bias term of each channel. We develop two simple methods for estimating the correction constants -one using iterative evaluation of the quantized network and one where the constants are set using a short training phase. Both methods are fast and require only a small amount of unlabeled data, making them appealing for rapid deployment of neural networks. Using the above methods we are able to match the performance of trainingbased quantization of MobileNets at a fraction of the cost.

No abstract