教導電腦識別圖片是一件多么具有顛覆性意義的事情。然而今早，透過谷歌研究團隊發布的JAMA論文，我們看到了這一技術的雛形。這支團隊致力于利用卷積神經網絡識別人眼視網膜的顯微照片。

瓦倫·古爾山、莉莉·彭和他們的同事使用了深度學習算法，研究了128,175幅美國和印度病患的視網膜圖像，隨后，54名在美國執業的眼科醫生對這些圖像進行了糖尿病視網膜病變鑒定。糖尿病視網膜病變的病理在于，連接眼睛后部（視網膜）的光敏器官中的微小血管出現壞死。長期的高血糖會損傷血管，導致其出血或滲血。這一病變將引發視覺模糊，并可能造成失明。該風險深深困擾著全球4.15億名糖尿病患者。

加州大學舊金山分校臨床醫學科學家、JAMA論文的通訊作者莉莉·彭說：“至少有三名眼科醫師對近13萬張圖片進行了評級，如果病例比較特殊的話，有時候最多會有7名醫師參與。然后，我們根據這些級別和這些圖片來培訓算法?！比缓?，該團隊測試了模型對兩個視網膜掃描（共計11,711幅圖片）臨床驗證組的糖尿病視網膜病變進行識別和適當評級的能力。這些圖片都已獲得了眼科專家的專業鑒定。

總體來說，谷歌算法以較高的敏感度和特異性，檢測出了測試圖片的糖尿病視網膜病變。莉莉說：“我們基本上證明了該技術的能力與獲美國職業認定、給驗證組評級的眼科醫生的能力旗鼓相當?！?

這項實驗的重要性何在？糖尿病視網膜病變在早期是可以預防的，但相對來講，全球各地很少有人能夠有機會進行專家篩選。谷歌的算法彌補了這一空白。它能夠可靠地在任何地方發揮作用，也可以在任何可工作的平臺上使用，例如智能手機或平板。

莉莉表示，“盡管可能需要龐大的計算集群來對這一模型進行培訓，但培訓之后的模型尺寸并不是那么大，它甚至可以裝入移動設備。”事實上，這一點也是谷歌團隊正在與印度的一些醫院共同解決的問題之一。

誰知道呢？也許在下一代的智能手機上，糖尿病患者就能夠掃描自己的眼睛，并獲得早期的預警信號。 （財富中文網）

?

作者：Clinton Leaf

譯者：馮豐

審校：夏林

How transformative can it be when you teach a computer to read images? Well, we’re getting an early glimpse of that this morning with the release of a JAMA paper by a team of Google researchers who trained a deep convolutional neural network to read photomicroscopic images of the backs of human eyes.

Varun Gulshan, Lily Peng, and colleagues used a deep learning algorithm to study 128,175 retinal images drawn from patients in the U.S. and India that were later reviewed for diabetic retinopathy (DR) by a group of 54 U.S.-licensed ophthalmologists. DR is a condition in which the tiny blood vessels in the light-sensitive tissue that lines the back of the eye (the retina) deteriorate. Chronic high blood sugar can damage the vessels, causing them to bleed or leak fluid, which distorts vision and can lead to blindness—a risk of profound concern to 415 million people with diabetes around the world.

“The nearly 130,000 images in this development set were graded by at least three ophthalmologists—sometimes up to seven if it was a tricky case—and then we trained an algorithm based upon those grades and those images,” says Google’s Lily Peng, a physician scientist trained at UCSF, who is the corresponding author on the JAMA paper. Then the team tested the model’s ability to identity and properly grade DR on two “clinical validation sets” of retinal scans (11,711 images in all) that had already been expertly characterized by eye specialists.

Overall, the Google algorithm detected DR on the test images with both high sensitivity and specificity. “We basically showed that we are on par with U.S. board-certified ophthalmologists who had graded the validation sets,” says Peng.

Why is this important? Diabetic retinopathy can be prevented if caught early—but relatively few people around the world have access to expert screening. That’s where Google’s algorithm comes in. It can conceivably be put to use anywhere—or anywhere a smartphone or tablet can work.

“While it may take acres of computer farms to actually train the model,” says Peng, “the model itself—once trained—is actually not that big, and can fit on even a mobile device.” That, in fact, is one of the things the Google team is now working on—in concert with some hospitals in India.

Who knows? Maybe in the next generation of smartphones, diabetics will be able to scan their own eyes for an early warning sign.